WO2023155532A1 - 位姿检测方法及装置、电子设备和存储介质 - Google Patents

位姿检测方法及装置、电子设备和存储介质 Download PDF

Info

Publication number
WO2023155532A1
WO2023155532A1 PCT/CN2022/134852 CN2022134852W WO2023155532A1 WO 2023155532 A1 WO2023155532 A1 WO 2023155532A1 CN 2022134852 W CN2022134852 W CN 2022134852W WO 2023155532 A1 WO2023155532 A1 WO 2023155532A1
Authority
WO
WIPO (PCT)
Prior art keywords
point cloud
image
frame
preset
target object
Prior art date
Application number
PCT/CN2022/134852
Other languages
English (en)
French (fr)
Inventor
王坤
林纯泽
王权
钱晨
Original Assignee
上海商汤智能科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海商汤智能科技有限公司 filed Critical 上海商汤智能科技有限公司
Publication of WO2023155532A1 publication Critical patent/WO2023155532A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
    • G06T7/75—Determining position or orientation of objects or cameras using feature-based methods involving models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/80—Analysis of captured images to determine intrinsic or extrinsic camera parameters, i.e. camera calibration
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10028—Range image; Depth image; 3D point clouds
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30196—Human being; Person
    • G06T2207/30201—Face
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30244—Camera pose

Definitions

  • the present disclosure relates to the field of computer technology, and in particular to a pose detection method and device, electronic equipment, and a storage medium.
  • Image-based 3D reconstruction of targets has become a very popular research direction, and has been widely used in target detection and recognition, target drive, augmented reality, etc.
  • the relative pose (Pose) has also formed a relatively mature logic scheme.
  • the present disclosure proposes a pose detection technical solution.
  • a pose detection method including: acquiring at least one frame of the first image, wherein the at least one frame of the first image includes at least one frame of the back of the target object captured by the image acquisition device. three-dimensional image; according to the at least one frame of the first image and the preset point cloud, determine the pose information of the target object in each of the at least one frame of the first image, wherein the preset point cloud is The 3D point cloud of the front face of the target object.
  • the preset point cloud includes N preset key points, where N is an integer greater than 1, and when the image acquisition device is a single camera, the at least one frame of the first image includes The single frame of the first image captured by the single camera; the determination of the pose information of the target object in the at least one frame of the first image according to the at least one frame of the first image and the preset point cloud, including : In the target object area of the single-frame first image, generate annotation information including N first key points; project the preset point cloud onto the single-frame first image, and determine N first key points The corresponding N first projection points; performing fitting processing on the N first projection points and the N preset key points to obtain pose information of the target object in the single frame of the first image.
  • the at least one frame of the first image includes: multiple frames of the first image and multiple frames of the second image, and the multiple frames of the second image are Multiple frames of two-dimensional images on the front of the target object; according to the at least one frame of the first image and the preset point cloud, determining the pose information of the target object in each of the at least one frame of the first image,
  • the method includes: performing three-dimensional reconstruction on the multiple frames of second images to obtain the preset point cloud, and performing three-dimensional reconstruction on the multiple frames of first images to obtain the first point cloud; according to the preset point cloud and the preset point cloud The relative relationship of the first point cloud is determined, and the pose information of the target object in each frame of the first image of the multiple frames of the first image is determined.
  • the image acquisition device when it is a camera array, before acquiring at least one frame of the first image, it includes: calibrating the camera array according to the checkerboard image acquired by the camera array internal parameters and external parameters, the internal parameters include at least one of focal length or pixels, and the external parameters include at least one of camera position or camera rotation angle; the three-dimensional reconstruction of the multiple frames of second images, Obtaining the preset point cloud includes: performing three-dimensional reconstruction on the multiple frames of second images according to the internal parameters and the external parameters to obtain the preset point cloud; and, the pairing of the multiple frames Performing three-dimensional reconstruction on the first image to obtain a first point cloud includes: performing three-dimensional reconstruction on the multiple frames of first images according to the internal parameters and the external parameters to obtain the first point cloud.
  • the relative relationship between the preset point cloud and the first point cloud includes pose transformation information between the preset point cloud and the first point cloud
  • the relative relationship between the preset point cloud and the first point cloud, and determining the pose information of the target object in each frame of the first image of the multiple frames of the first image includes: according to the content of the camera array parameters and external parameters, and pose transformation information between the preset point cloud and the first point cloud, determine the relationship between the preset point cloud and each frame of the first image in the multiple frames of first images
  • the pose transformation information according to the pose transformation information between the preset point cloud and each frame of the first image in the multi-frame first image, determine the first image in each frame of the multi-frame first image The pose information of the target object.
  • the relative relationship between the preset point cloud and the first point cloud includes pose transformation information between the preset point cloud and the first point cloud
  • the method It also includes: performing point cloud registration processing on the preset point cloud, determining grid data corresponding to the preset point cloud, and pose transformation information between the grid data and the preset point cloud;
  • the determining, according to the relative relationship between the preset point cloud and the first point cloud, the pose information of the target object in each frame of the first image of the multiple frames includes: according to the grid data and the pose transformation information between the preset point cloud and the pose transformation information between the preset point cloud and the first point cloud, determine the grid data and the first point cloud According to the internal parameters and external parameters of the camera array, the pose transformation information between the grid data and the first point cloud, determine the grid data and the multiple The pose transformation information between each frame of the first image in the frame first image; according to the pose transformation information between the grid data and each frame of the first image in the multiple frames of the first image, determine the multiple The pose information of the target object in
  • the method further includes: performing alignment processing on the preset point cloud and the first point cloud, and determining a relative relationship between the preset point cloud and the first point cloud.
  • the aligning the preset point cloud with the first point cloud, and determining the relative relationship between the preset point cloud and the first point cloud include : In the face area of the preset point cloud, generate labeling information including N preset key points; in the face area of the first point cloud, generate labeling information including K second key points , K is an integer not greater than N; align the N key points of the preset point cloud with the K second key points of the first point cloud, and determine the preset point cloud and the first point Relative relationship between clouds.
  • the target object includes body parts of the human body; when the body parts of the human body include a human face, the preset key points include key points of human face organs; When the body parts include limbs, the preset key points include limb key points.
  • a pose detection device including: an acquisition module, configured to acquire at least one frame of a first image, where the first image includes a two-dimensional image of the back of a target object captured by an image acquisition device; A determining module, configured to determine pose information of the target object in the first image according to the at least one frame of the first image and a preset point cloud, wherein the preset point cloud is a three-dimensional point on the front of the target object cloud.
  • an electronic device including: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to call the instructions stored in the memory to execute the above-mentioned method.
  • a computer-readable storage medium on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the above method is implemented.
  • At least one frame of the first image to be processed can be acquired, the first image includes a two-dimensional image of the back of the target object captured by the image acquisition device, based on the at least one frame of the first image, and the preset target
  • the three-dimensional point cloud of the front of the object determines the pose information of the target object in the first image.
  • the pose of the multi-frame images captured by the device effectively solves the all-round pose detection of the target object, regardless of the angle between the face orientation of the target object and the shooting direction, it can achieve accurate pose detection.
  • Fig. 1 shows a flowchart of a pose detection method according to an embodiment of the present disclosure.
  • Fig. 2 shows a schematic diagram of a preset key point and a first key point according to an embodiment of the present disclosure.
  • FIG. 3 shows a schematic diagram of a projection effect according to an embodiment of the present disclosure.
  • FIG. 4 shows a schematic diagram of a calibration plate according to an embodiment of the disclosure.
  • Fig. 5 shows a schematic diagram of a preset point cloud acquisition method according to an embodiment of the present disclosure.
  • Fig. 6 shows a schematic diagram of a first point cloud acquisition method according to an embodiment of the present disclosure.
  • Fig. 7 shows a schematic diagram of mesh data and first point cloud alignment according to an embodiment of the present disclosure.
  • FIG. 8 shows a schematic diagram of a projection effect according to an embodiment of the present disclosure.
  • FIG. 9 shows a block diagram of a pose detection device according to an embodiment of the present disclosure.
  • Fig. 10 shows a block diagram of an electronic device according to an embodiment of the present disclosure.
  • Fig. 11 shows a block diagram of an electronic device according to another embodiment of the present disclosure.
  • the pose detection of the target object can be widely used in various scenarios, where the target object can include body parts of the human body, etc., for example, the target object has clear semantic parts (such as eye corners, mouth corners, nose tips, etc.) Objects that define semantic parts.
  • the target objects of the embodiments of the present disclosure are not limited to body parts of the human body (such as human heads, limbs, etc.), and any object that has clear semantic parts on the front and lacks clear semantic parts on the back can be used as the target object.
  • the subsequent disclosed embodiments are described by taking the target object as a human head as an example.
  • the case where the target object is other objects can be flexibly expanded according to the subsequent disclosed embodiments, which is not limited in the present disclosure.
  • the pose detection of the human head for the face/head 360-degree reconstruction task scene, in order to improve the accuracy of the face/head 360-degree reconstruction, it is necessary to perform pose detection on the head image at any angle within the 360-degree range. Determine the pose of the face/head with respect to the front and back images to provide the training data required for this 360° reconstruction task. Or, in some scenes that need to use the head orientation as an important judgment basis, it is also necessary to detect the pose of the head in the head image. For example, in the detection scene of safe driving, if the vehicle camera is set behind the driver's The image of the back of the driver's head determines whether the driver is looking left and right.
  • the application scenario of Augmented Reality (AR) interaction based on face pose it is necessary to perform pose detection on the captured face image.
  • AR Augmented Reality
  • the application scenario of augmented reality AR interaction based on face pose may include, in addition to special effects of face stickers, virtual makeup processing, special effects of human face conversion to animal faces, etc., which is not limited in this disclosure.
  • the pose detection of the target object in the image to be processed can estimate the orientation or attitude of the target object relative to a certain reference benchmark, where the reference benchmark can be a preset three-dimensional model of the target object, and in actual use, it can be based on the actual It is necessary to select the reference datum of the target object pose.
  • the reference benchmark can be a preset three-dimensional model of the target object, and in actual use, it can be based on the actual It is necessary to select the reference datum of the target object pose.
  • the pose detection of the target object in the image to be processed mainly focuses on the pose detection of the target object facing the camera and rotating within 90 degrees left and right, that is, the face orientation of the target object and the shooting direction (face The angle between the direction and the image acquisition device) is between 0° to 90°, and between 270° to 360°.
  • face orientation of the target object and the shooting direction face The angle between the direction and the image acquisition device
  • the screened image of the target object has clear semantic parts (such as the corners of the eyes, the corners of the mouth, the tip of the nose, etc.), and the key points can be extracted from the screened images of the target object, and the 3D model of the target object can be obtained based on the key point fitting, and then The three-dimensional model can be used as a reference to determine the pose of the target object in the image relative to the three-dimensional model.
  • the 3D model can be a standardized model, such as a 3D dense point cloud.
  • the standardized model can have a topology structure, that is, the connection relationship between each point in the model is determined, and each point in the model and each part of the target object are also bound.
  • the pose detection method mentioned above needs to extract the key points of the target object in the image, and fit the 3D model of the target object based on the key points, so as to determine the pose of the target object in the image relative to the 3D model.
  • the angle between the face orientation of the target object and the shooting direction is between 0° and 90°, and between 270° and 360°
  • a target object with a small rotation angle for example, a front face image
  • the target object in the image With clear semantic parts (such as eye corners, mouth corners, nose tip, etc.) and many key points, it can accurately fit the 3D model and estimate the pose of the target object relative to the 3D model.
  • the target object with a relatively large rotation angle for example, the image of the back of the head
  • the target object area in the image includes only a few
  • the clear semantic parts such as the corners of the eyes, the corners of the mouth, the tip of the nose, etc.
  • even some target object images do not include clear semantic parts at all, and the number of key points that can be extracted is small or even none. model, resulting in a large error in the pose detection results of the target object relative to the 3D model. It can be seen that the related technology cannot effectively solve the accurate pose detection of the target object at any shooting angle, especially cannot effectively solve the accurate pose detection when the back of the target object is at the shooting angle.
  • the present disclosure provides a pose detection technical solution, which can acquire at least one frame of the first image to be processed, the first image includes the two-dimensional image of the back of the target object collected by the image acquisition device, according to at least one frame
  • the first image as well as the preset 3D point cloud on the front of the target object, determine the pose information of the target object in the first image.
  • the pose of a single frame image taken with the back facing the camera it can also accurately realize the pose of multi-frame images taken by the camera array equipment, effectively solve the all-round pose detection of the target object, no matter what angle the face orientation of the target object presents and the angle of the shooting direction, the accurate position can be achieved posture detection.
  • Fig. 1 shows a flow chart of a pose detection method according to an embodiment of the present disclosure
  • the pose detection method may be executed by an electronic device such as a terminal device or a server
  • the terminal device may be a user equipment (User Equipment, UE), a mobile device, User terminal, cellular phone, cordless phone, personal digital assistant (Personal Digital Assistant, PDA), handheld device, computing device, vehicle-mounted device, wearable device, etc.
  • the method can call the computer-readable instructions stored in the memory through the processor way to achieve.
  • the method may be performed by a server.
  • the pose detection method of the embodiment of the present disclosure will be described below by taking an electronic device as an execution subject as an example.
  • the pose detection method includes:
  • step S1 at least one frame of the first image is acquired, wherein the at least one frame of the first image includes at least one frame of two-dimensional image of the back of the target object collected by the image acquisition device;
  • the target object may include objects that have clear semantic parts (for example, including eye corners, mouth corners, nose tip, etc.) on the front (for example, including the face), but lack clear semantic parts on the back (for example, including the back of the head).
  • the target object of the embodiments of the present disclosure is not limited to the head, and any object that has a clear semantic part on the front and lacks a clear semantic part on the back can be used as the target object.
  • Subsequent disclosed embodiments are described by taking the target object as a human head as an example. The situation that the target object is other objects can be flexibly expanded according to the subsequent disclosed embodiments, which is not limited in the present disclosure.
  • the first image to be processed may be a pre-acquired image including the target object, or an image acquired when the target object is photographed by an image acquisition device.
  • the electronic device is a terminal device such as a mobile phone or a computer
  • the first image including the target object may be collected by an image acquisition device connected to the terminal device, or the first image including the target object may be selected from the photo album of the terminal device , or receive the first image including the target object from other application programs installed in the terminal device.
  • the image acquisition device may be a single camera, or may be a camera array composed of multiple cameras at different spatial positions.
  • one frame or multiple frames of the first image may be acquired in one shot. If the image acquisition device is a single camera, under the condition that the relative position and angle between the single camera and the target object do not change, a frame of the first image to be processed can be obtained, and the first image is the image of the target object corresponding to this shooting angle. 2D image.
  • the image acquisition device is a camera array, for example, for a camera array composed of M (M is an integer greater than 1) cameras at different spatial positions, under the condition that the relative position and angle between the camera array and the target object do not change, it can be M frames of first images to be processed are acquired, and each frame of first images corresponds to a two-dimensional image of a target object at a shooting angle of a respective camera.
  • the image acquisition device in the embodiment of the present disclosure may not only be a camera, but may also be a device with a camera function such as a camera, a video camera, a scanner, a mobile phone, and a tablet computer, which is not limited in the present disclosure.
  • the image acquisition device can acquire images of the target object at various angles. For example, when the target object is facing the image acquisition device, that is, the angle between the face orientation of the target object and the shooting direction is between 0° to 90°, and 270° to 360°, the to-be-processed information about the front of the target object can be collected. or at least one frame of the first image; or, in the case where the target object is facing away from the image acquisition device, that is, the angle between the face orientation of the target object and the shooting direction is between 90 degrees and 270 degrees, the image of the back of the target object can be collected At least one frame of the first image to be processed.
  • the pose detection method in the embodiment of the present disclosure can be applied not only to the front image of the target object, but also to the image on the back of the target object , perform pose detection on a 2D image of the back of the target object.
  • the embodiments of the present disclosure focus on introducing a method of performing pose detection of the target object in at least one frame of the acquired first image to be processed when the first image is a two-dimensional image of the back of the target object.
  • step S2 according to the at least one frame of the first image and the preset point cloud, the pose information of the target object in each of the at least one frame of the first image is determined.
  • the preset point cloud is a three-dimensional point cloud of the front of the target object
  • the pose information includes at least one of the following information: scaling information, rotation information and translation information.
  • the scaling information may represent the scale of enlargement or reduction between the target object and the 3D point cloud in the first image
  • the rotation information may represent the rotation angle between the target object and the 3D point cloud in the first image
  • the translation information may Indicates the direction and distance of movement between the target object in the first image and the 3D point cloud.
  • the target object in the first image is two-dimensional data
  • the three-dimensional point cloud is three-dimensional data
  • the target object in the first image can be calculated Regarding the scaling information, rotation information, and/or translation information between the object and the two-dimensional projection of the three-dimensional point cloud
  • this disclosure does not make any specific determination of the scaling information, rotation information, and/or translation information between specific two-dimensional data and three-dimensional data. limit.
  • the determined pose information of the target object can be applied to interactive entertainment scenes based on augmented reality AR such as film and television, games, and virtual social interaction.
  • the pose information of the target object has It is beneficial to judge the relative position between the target object and virtual information (such as virtual scenes, virtual objects, etc.) in reality, and realize the combination or interaction between the target object and virtual information.
  • the pose of the target object is used Information, can realize that multiple virtual fruits fly to the target object at different angles.
  • the pose information of the target object can be used to adjust the display angle and position of the virtual clothing to improve the dressing effect.
  • the pose information of the target object is conducive to the combination of virtual special effects and target objects, such as virtual makeup processing, human face conversion animal face special effects, human face sticker special effects, etc.
  • target objects such as virtual makeup processing, human face conversion animal face special effects, human face sticker special effects, etc.
  • the present disclosure does not limit the application scenarios of the pose information of the target object.
  • the preset point cloud is a three-dimensional point cloud used to describe the front of the target object, for example, including the three-dimensional point cloud of the front of the target object constructed under the world coordinate system (World Coordinate System), the preset The point cloud can be used as a reference for object pose estimation.
  • World Coordinate System World Coordinate System
  • the pre-stored preset point cloud can be read from the memory of the electronic device, and the multi-frame frontal two-dimensional image of the target object can also be three-dimensionally reconstructed to obtain the preset point cloud of the target object.
  • the preset point cloud is obtained by performing surveying and mapping, and the present disclosure does not limit the method of determining the preset point cloud.
  • the corresponding relationship between the image coordinate points of the target object in at least one frame of the first image and the three-dimensional point cloud (preset point cloud) on the front of the target object can be analyzed, and at least one frame of the first image can be determined
  • Pose information of the target object, the pose information may include at least one of scaling information, rotation information, and translation information of the target object in the first image relative to a preset point cloud.
  • the projection method can be used to The three-dimensional preset point cloud is projected into the two-dimensional image space, and the corresponding relationship between the first image and the projected preset point cloud is analyzed, and then the pose information of the target object in at least one frame of the first image is determined; or, it is also possible
  • the projection method can be used to The three-dimensional preset point cloud, perform three-dimensional reconstruction on at least one frame of the first image, so as to reconstruct the three-dimensional point cloud (ie, the first point cloud) of the target object, and then analyze the reconstructed first point cloud and the preset point The corresponding relationship of the cloud, and then determine the pose information of the target object in at least one frame of the first image.
  • multiple preset key points can be marked at the preset point cloud with clear semantic parts, and the first frame reconstructed on at least one frame of the first image or based on at least one frame of the first image
  • the key points with the same point sequence as the preset key points are marked on the point cloud.
  • the corresponding relationship between the key points of the preset point cloud and the key points on the at least one frame of the first image or the reconstructed first point cloud based on the at least one frame of the first image can be analyzed, and then at least one frame of the first image can be determined pose information of the target object.
  • the preset key point of the preset point cloud can be determined by analyzing the corresponding relationship between the preset key point of the preset point cloud and the first key point of the single frame of the first image.
  • the first projection point of the key point is fitted with the first key point of the single frame of the first image to obtain the pose information of the target object in the single frame of the first image, and the pose information includes at least one of the following information : Zoom information, rotation information and translation information.
  • the scaling information may indicate the ratio of enlargement or reduction between the target object in the single-frame first image and the projection of the preset point cloud
  • the rotation information may indicate the ratio between the target object in the single-frame first image and the projection of the preset point cloud.
  • the translation information may represent the moving direction and distance between the target object and the projection of the preset point cloud in the single frame of the first image.
  • the first point cloud corresponding to the target object on the back can be constructed through the 3D point cloud reconstruction technology, and the second key point can be marked on the first point cloud, and the preset The key point is aligned with the second key point, that is, the preset point cloud is aligned with the first point cloud, so as to obtain the pose transformation information between the preset point cloud and the first point cloud, and then according to the pose transformation information, Determine pose information of a target object in multiple frames of first images.
  • the pose information includes at least one of the following information: scaling information, rotation information and translation information.
  • the scaling information may indicate the ratio of enlargement or reduction between the projection of the target object and the preset point cloud in the first image of each frame
  • the rotation information may indicate the ratio between the projection of the target object and the preset point cloud in the first image of each frame.
  • the translation information may indicate the moving direction and distance between the target object and the projection of the preset point cloud in each frame of the first image.
  • steps S1-S2 for some back images of target objects that cannot accurately provide clear semantic parts, it is also possible to accurately determine the pose information of the target object in the image relative to the 3D model (preset point cloud) of the target object, for example.
  • steps S1-S2 not only can the pose of the target object be accurately determined in the single-frame image taken with the camera facing away from the camera, but also the pose of the target object in the multi-frame images taken with the camera array device facing away can be accurately determined, effectively realizing the All-round pose detection, regardless of the angle between the face orientation of the target object and the shooting direction, the pose detection can be accurately realized.
  • the poses of the embodiments of the present disclosure are respectively analyzed.
  • the detection method is illustrated as an example.
  • step S1 if the image acquisition device is a single camera, at least one frame of the first image to be processed may be acquired, and the at least one frame of the first image includes the single camera Single frame first image captured.
  • the single frame of the first image may be a two-dimensional image taken when the person's head faces away from a single camera in a natural (In The Wild) scene.
  • the single frame of the first image may be the orientation of the face of the person and the shooting direction An image at any angle between 90 degrees and 270 degrees. It should be understood that 90 degrees to 270 degrees is only used as a reference range, and embodiments of the present disclosure are not limited to the above numerical ranges.
  • the natural scene is a real scene in the real world (such as work, study, life), for example, it may include a single-frame first image taken in a park, office, school, shopping mall, etc., a single-frame image taken in a natural scene
  • the background of the first image can be varied.
  • step S1 a single frame of the first image captured by a single camera can be obtained.
  • step S2 according to the first key point of the single frame of the first image and the preset key points of the preset point cloud, the human head in the first image can be determined. pose information, step S2 may include:
  • step SA1 in the target object area of the single frame of the first image, generate annotation information including N first key points, and the annotation information may include position information of the first key points;
  • step SA2 the preset point cloud is projected onto the single frame of the first image, and N first projection points corresponding to the N first key points are determined;
  • step SA3 fitting processing is performed on the N first projected points and the N preset key points to obtain pose information of the target object in the single frame of the first image.
  • FIG. 2 shows a schematic diagram of a preset key point and a first key point according to an embodiment of the present disclosure.
  • the 3D model of the human head in the left figure is a preset point cloud (i.e.
  • the 3D point cloud on the front of the human head which can include N preset key points, for example, the number of preset key points
  • the preset key points include the key points of human face organs
  • the key points include: left/right outer corner of the eye, left/right inner corner of the eye, tip of the nose, left/right The mouth corner point, the center point of the chin, and the left/right ear base point, among which, the order of the number labels in the left picture of Figure 2 represents the point sequence of the preset key points.
  • the target object may include body parts of the human body.
  • the preset key points include key points of human face organs;
  • the preset key points include limb key points, and the present disclosure does not specifically limit the preset key points.
  • the image on the back of the human head on the right side is the first single-frame image taken by a single camera acquired in step S1.
  • the first single-frame image can be The head area of the head area generates annotation information containing N first key points.
  • N first key points can be marked in the head area of the first single frame image according to the received user operation information, and N first key points can be generated including N first key points.
  • Annotation information of key points a preset neural network model can also be used to input a single frame of the first image into the neural network model to generate annotation information containing N first key points, wherein the preset neural network model is also After the neural network is trained (learned) through a large number of samples, a neural network model capable of marking the first key point based on the head area of the first single frame image is obtained.
  • the present disclosure does not limit the method for generating annotation information.
  • the marked positions of the N first key points on the back of the human head may correspond to the marked positions of the N preset key points, and have the same point sequence as the N preset key points.
  • the first key points may correspond to the left/right outer corner of the eye, the left/right inner corner of the eye, the tip of the nose, the left/right corner of the mouth,
  • the center point of the chin, the left/right ear root point, and the order of the numbers in the right figure of Figure 2 represent the point sequence of the first key point.
  • the ear part can also be seen in the back image of the human head, in order to estimate the size of the face and improve the accuracy of pose detection, when selecting the first key point to be marked and the preset key point, at least the left/ Right ear point.
  • the greater the value of the number N of the first key point and the preset key point the greater the calculation amount of the pose detection method in the embodiment of the present disclosure, and the more accurate the corresponding pose detection will be.
  • the present disclosure does not limit the specific value of the number N of annotations of the first key point, and the location of the annotations.
  • the three-dimensional preview shown in the left figure of Figure 2 can be The point cloud is projected to the first two-dimensional single-frame image as shown in the right figure of Figure 2, so that the first projection points of the N preset key points on the preset point cloud and the N first key points in the single-frame first image point-phase fitting to obtain the pose information of the target object corresponding to the single-frame first image.
  • P3d represents N preset key points in the preset point cloud, namely: (x 1 , y 1 , z 1 ) ⁇ (x N , y N , z N ), and P2d represents The N first key points of , namely: (x 1 ′, y 1 ′) ⁇ (x N ′, y N ′).
  • the equation can be constructed based on projection, so that the first projection point of the N preset key points P3d on the preset point cloud and the N first key points in the first image of the single frame P2d phase fitting, namely:
  • P2d Project(scale*rotation*P3d+translation) (1).
  • Project represents the projection function, which can be used to reduce the dimension of the target object, and can project the target object in the three-dimensional space to a two-dimensional plane.
  • the specific projection methods can include orthogonal projection, perspective projection, etc.
  • the present disclosure does not limit the types of projection methods.
  • the pose transformation can be performed on the N key points p3d of the preset point cloud, for example, it can include scaling transformation scale, rotation Transform the rotation and translation transformations, and use the value of the pose transformation that makes the equation in formula (1) to be established as the pose information of the target object in the first image of the single frame.
  • the pose information can include solving the formula (1) The obtained zoom information scale, rotation information rotation, and translation information translation.
  • the scaling transformation scale may be a parameter greater than 0, the rotation transformation rotation may be a 3*3 matrix, and the translation transformation may be a 2*1 matrix.
  • the matrix parameters related to the rotation transformation and the translation transformation are not limited to the above examples.
  • the N first projection points corresponding to the N first key points are determined; based on fitting processing methods such as the least square method, the N The first projection point is fitted to the N preset key points to obtain pose information of the target object in the single frame of the first image.
  • the preset point cloud can be projected onto the single frame of the first image to determine N first key points p3d: (x 1 , y 1 , z 1 ) ⁇ (x N , y N , z N )
  • the corresponding N first projection points are: Project(scale*rotation*P3d+translation), that is: (px 1 , py 1 ) ⁇ (px N , py N );
  • a linear optimization method such as the least square method, can be used to solve the formula (1) to obtain the pose information of the target object in the first image of the single frame, namely:
  • (xi ′ , y i ′) represents the first key point
  • ( pxi , py i ) represents the first projected point, which will make the value of formula (2) the smallest scale information scale
  • rotation information are determined as the pose information of the target object in the single frame first image.
  • label information containing N first key points can be generated on the back of the head, so that the N first key points and the frontal face model (preset The N first projected points of the point cloud) have the same point sequence, and then directly according to the fitting process of the paired N groups of points (for example, using the least squares method), the pose information of the first image of the single frame can be solved. It is easier and more efficient to obtain the pose information of the first single frame image.
  • the pose information of the target object in the single frame of the first image can also be accurately obtained for the single frame of the first image shot facing away from the camera.
  • steps SA1-SA3 can also be applied to the frontal image of the target object.
  • the pose information of the frontal image of the human face can be obtained, wherein the first key point can be interactively Manual labeling, feature point detection (for example, including based on deep learning methods, etc.) can also be performed on the face frontal image to obtain the first key point.
  • FIG. 3 shows a schematic diagram of a projection effect according to an embodiment of the present disclosure.
  • FIG. 3 takes a human head as a target object and respectively shows the projection effect of a single frame of a first image at two different shooting angles.
  • the 10 points of the triangle represent the 10 first key points marked in the first image of the single frame
  • the 10 points of the circle represent the three-dimensional preset key points (the 10 points in the left figure of Figure 2 ) is projected to the first projection point of the single frame of the first image.
  • the first key point marked in the first single frame image and the first projection point can coincide or approach each other, which shows that the pose detection method of the embodiment of the present disclosure can more accurately solve the pose of the image of the back of the head information.
  • the above describes the pose detection for the scene of a single frame of rear head images captured by a single camera.
  • the following describes the pose detection method of the embodiment of the present disclosure for the scene of multiple frames of rear head images captured by a camera array in an experimental scene.
  • the experimental scene is an artificially arranged scene, which may include an artificially arranged laboratory, so that the background of the image taken in this scene is artificially set, for example, the background is pure white.
  • step S1 if the image acquisition device is a camera array, at least one frame of the first image to be processed may be acquired, and the at least one frame of the first image may include: multiple frames The first image and multiple frames of second images, the second images are two-dimensional images of the front of the target object.
  • step S1 multiple frames of the first image and multiple frames of the second image are determined, and in step S2, each frame of the first image can be determined according to the multiple frames of the first image and the preset point cloud three-dimensionally reconstructed from the multiple frames of the second image
  • the pose information of the target object, step S2 may include:
  • step SB2 perform three-dimensional reconstruction on the multiple frames of second images to obtain the preset point cloud, and perform three-dimensional reconstruction on the multiple frames of first images to obtain a first point cloud;
  • step SB3 according to the relative relationship between the preset point cloud and the first point cloud, determine the pose information of the target object in each frame of the first image in the multiple frames of the first image.
  • camera arrays can use multiple cameras at different spatial positions to collect images from different perspectives.
  • the entire camera array can be regarded as a multiple-center-of-projection camera (Multiple-Center-of-Projection Camera), which can obtain multi-view information of the target object.
  • the camera array includes M cameras
  • M frames of images of different viewing angles can be collected for one shot of the target object, and the present disclosure does not limit the number M of cameras included in the camera array.
  • step S1 the same camera array may be used to capture multiple frames of first images and multiple frames of second images, or different camera arrays may be used to capture multiple frames of first images and multiple frames of second images respectively, and the acquired
  • the number of frames of the first image and the number of frames of the acquired second image may be the same or different, and the present disclosure does not limit the number of frames of the first image and the number of frames of the second image.
  • an array including M (for example, 30) cameras may be used to acquire M frames of first images and M frames of second images.
  • the image acquisition device is a camera array
  • the image acquisition device before acquiring M frames of the first image and M frames of the second image through the camera array (that is, before step S1), in step SB1, according to the checkerboard images collected by the camera array, internal parameters and external parameters of the camera array are calibrated, the internal parameters include at least one of focal length or pixels, and the external parameters include at least one of camera position or camera rotation angle.
  • the checkerboard image means an image with a checkerboard pattern
  • the checkerboard pattern is an image composed of two colors (eg, black and white) alternately.
  • a checkerboard image can be composed of 2 rows ⁇ 3 columns of squares, the first row, the first column, the first row, the third column, and the second row, the second column are black squares, the first row, the second column, the second The first column of the row and the second row and the third column are white squares, and the disclosure does not limit the number of rows, columns, colors, and side lengths of each checkerboard.
  • Calibrating the camera array can determine the relationship between the three-dimensional geometric position of a point on the surface of the target object in three-dimensional space and its corresponding point in the image, that is, the mapping from world coordinates to pixel coordinates.
  • the calibration parameters of the M cameras included in the camera array can be solved simultaneously to obtain the calibration parameter trans_camera of the camera array.
  • the calibration parameter trans_camera can include internal parameters and external parameters.
  • the external parameters are used to determine the relationship between the geometric position of a point on the surface of the target object in the three-dimensional space of the real world and its corresponding point in the camera coordinates, that is, the mapping from the world coordinates to the camera coordinates. It can include camera position, camera rotation angle, etc.; internal parameters are used to determine how a certain point on the surface of the target object under the camera coordinates continues to pass through the lens of the camera, and becomes a pixel in the image through pinhole imaging and electronic conversion, that is, the camera
  • the mapping from coordinates to image coordinates may include parameters related to the characteristics of the camera itself, such as focal length, pixels, and so on.
  • Calibration methods used to calibrate the camera array may include: traditional camera calibration methods, active vision camera calibration methods, camera self-calibration methods, zero-distortion camera calibration methods, etc.
  • the present disclosure does not limit the calibration methods.
  • the embodiments of the present disclosure can, for example, use a single-plane checkerboard camera calibration method, which can reduce or overcome the shortcomings of high-precision calibration objects required by traditional camera calibration methods, and only use one
  • the printed checkerboard can be used to calibrate the camera.
  • the single-plane checkerboard camera calibration method can improve the accuracy and is more convenient to operate.
  • the single-plane checkerboard camera calibration method can perform camera calibration based on a single-plane checkerboard.
  • the checkerboard used for calibration is a plane in a three-dimensional scene (world coordinates), and the two-dimensional image (image coordinates) captured by it is another plane. , according to the coordinates of the corresponding points of the two planes, the camera calibration parameter trans_camera can be solved.
  • the camera array consists of cameras 1 ⁇ cameras M
  • Its calibration parameter trans_camera[i] can include internal parameters composed of matrix A and external parameters composed of matrix [RT], assuming homogeneous coordinates in the three-dimensional world coordinate system Homogeneous coordinates in a two-dimensional image coordinate system
  • the homography from the checkerboard plane to the image plane for calibration can be expressed as:
  • s represents the scale factor from the world coordinate system to the image coordinate system.
  • ⁇ and ⁇ represent the x-axis and y-axis directions of each pixel in the image coordinate system
  • the physical size of , c represents the distortion parameter of the physical coordinates of the image, u 0 and v 0 represent the origin of the image; in the external parameters [RT], R represents the rotation transformation from the world coordinate system to the camera coordinate system, and T represents the transformation from the world coordinate system to The translation transformation of the camera coordinate system.
  • Fig. 4 shows a schematic diagram of a calibration board according to an embodiment of the present disclosure.
  • a single-plane checkerboard camera calibration method can be used to calculate the calibration parameter trans_camera of the camera using a checkerboard grid, wherein the calibration board can be 12 ⁇ 9 checkerboards, the size of each checkerboard can be a square with a side length of 10mm.
  • the checkerboard calibration board shown in FIG. 4 is only an example, and the present disclosure does not limit the number of rows and columns of checkerboard grids on the calibration board, and the size of each checkerboard grid.
  • the calibration parameters (internal parameters + external parameters) of the camera array can be obtained, which can be used to determine the relationship between the image coordinate system and the world coordinate system of each camera in the camera array.
  • the camera array can be used to collect images of the target object in step S1.
  • the camera array includes M cameras set at different angles, it is possible to obtain M frames of first images when the front of the target object (such as a human face) is facing away from the camera array, and when the front of the target object (such as a human face) is facing away from the camera array, and face) towards the camera array, acquire M frames of second images.
  • M frames of the first image and M frames of the second image are taken as an example, and the number of frames of the first image and the number of frames of the second image may also be different, which is not limited in the present disclosure.
  • step SB2 three-dimensional reconstruction may be performed on M frames of second images to obtain a preset point cloud, and three-dimensional reconstruction may be performed on the M frames of first images to obtain a first point cloud.
  • the preset point cloud is the three-dimensional point cloud of the front of the target object, which can be used as a reference for pose detection
  • the first point cloud is the three-dimensional point cloud of the back of the target object.
  • the method for three-dimensional reconstruction of the target object for multiple frames of images collected by the camera array may include: a reconstruction method based on traditional methods such as binocular vision, a reconstruction method based on models such as a three-dimensional deformable model (3D Morphable Model), a reconstruction method based on The reconstruction method of convolutional neural network, the reconstruction method based on 3D modeling software, etc., the embodiments of the present disclosure do not limit the specific 3D reconstruction method.
  • step SB2 may include: performing three-dimensional reconstruction on the multiple frames of second images according to the internal parameters and the external parameters to obtain the preset point cloud, and parameters and the extrinsic parameters, performing three-dimensional reconstruction on the multiple frames of first images to obtain the first point cloud.
  • a preset 3D reconstruction network may be used to input the internal parameters, the external parameters, and the M frames of second images into the 3D reconstruction network to generate a preset point cloud, and the internal parameters and the The external parameters and the M frames of first images are input to the 3D reconstruction network to generate a preset point cloud.
  • the preset 3D reconstruction network is also a trained neural network. After a large number of sample training (learning), the 3D reconstruction network can generate 3D point clouds based on input internal parameters, external parameters, and multi-frame images. There are no restrictions on the network structure and network training method of the 3D reconstruction network.
  • the 3D modeling software can be implemented through related technologies, and the present disclosure does not limit the 3D modeling software.
  • the calibration parameters (internal parameters + external parameters) of the camera array obtained in step SB1 and the M obtained in step SB1 can be
  • FIG. 6 shows a schematic diagram of a first point cloud acquisition method according to an embodiment of the present disclosure.
  • the calibration parameters internal parameters+external parameters
  • the back three-dimensional dense point cloud shown on the right is the first point cloud part_scan.
  • step SB2 the preset point cloud whole_scan and the first point cloud part_scan are obtained, and in step SB3, the pose of the target object in the first image of each frame can be determined according to the relative relationship between the preset point cloud whole_scan and the first point cloud part_scan information.
  • the relative relationship between the preset point cloud and the first point cloud includes pose transformation information between the preset point cloud and the first point cloud.
  • an alignment process may be performed on the preset point cloud whole_scan and the first point cloud part_scan to determine a relative relationship between the preset point cloud and the first point cloud.
  • the initial pose of the preset point cloud whole_scan can be determined according to the marked preset key points, and the initial pose of the first point cloud part_scan can be determined according to the second key point.
  • the preset point cloud whole_scan can be aligned with the first point cloud part_scan based on the initial pose of the preset point cloud whole_scan and the initial pose of the first point cloud part_scan.
  • the preset point cloud whole_scan can be kept When the initial pose remains unchanged, change (for example, at least one of rotation, scaling, and translation) the pose of the first point cloud part_scan, so that the preset point cloud whole_scan is aligned with the first point cloud part_scan; or, it is also possible While keeping the initial pose of the first point cloud part_scan unchanged, change the pose of the preset point cloud whole_scan so that the pose of the preset point cloud whole_scan is aligned with the initial pose of the first point cloud part_scan; or, It is also possible to change the initial pose of the preset point cloud whole_scan and the initial pose of the first point cloud part_scan at the same time to align the two. This disclosure does not limit this.
  • the preset point cloud whole_scan and the first point cloud part_scan can be aligned, and the iteration closest points (ICP) method can be used to determine the pose transformation of the preset point cloud whole_scan and the first point cloud part_scan info trans_scan.
  • ICP iteration closest points
  • trans_scan scale_2*rotation_2*whole_scan +translation_2.
  • the difference between the pose transformation information trans_scan of the preset point cloud whole_scan and the first point cloud part_scan is the smallest (the pose transformation information trans_scan of the preset point cloud whole_scan is infinitely close to the first point cloud part_scan, that is, the preset point cloud Whole_scan is aligned with the first point cloud part_scan, as shown in the figure in the middle of Figure 7), and the pose transformation information trans_scan of the preset point cloud whole_scan and the first point cloud part_scan is recorded.
  • step SB3 After determining the relative relationship between the preset point cloud whole_scan and the first point cloud part_scan, in step SB3, according to the relative relationship between the preset point cloud whole_scan and the first point cloud part_scan, determine the target in the first image of each frame The pose information of the object.
  • step SB3 may include:
  • step SB31' determine the pose transformation information between the preset point cloud and the first point cloud according to the internal parameters and external parameters of the camera array, and determine the relationship between the preset point cloud and the first point cloud of each frame. Pose transformation information between images;
  • step SB32' according to the pose transformation information between the preset point cloud and each frame of the first image, determine the pose information of the target object in each frame of the first image in the multiple frames of the first image.
  • step SB31' according to the calibration parameter trans_camera of the camera array obtained in step SB1, including the internal parameters and external parameters of the camera array, and the pose transformation information trans_scan of the preset point cloud whole_scan and the first point cloud part_scan , the pose transformation information of the preset point cloud whole_scan relative to the M frames of the first image can be determined.
  • the preset point cloud whole_scan can be respectively projected to each first image according to the pose transformation information of the preset point cloud whole_scan relative to the M frames of the first image, and it is determined that each frame of the first image corresponds to the target object pose information
  • the pose information may include at least one of scaling information, rotation information, and translation information.
  • the pose information of the target object in each frame of the first image in the multiple frames of the first image can be determined, and the method is simple and convenient.
  • the preset point cloud whole_scan is a messy point sequence (without regular topology) and has noise, and the pose of the target object in each frame of the first image of the multi-frame first image is directly determined according to the preset point cloud whole_scan information, it is easy to reduce the accuracy of pose detection.
  • the preset point cloud whole_scan can be unified into grid data (mesh data) with regular topology to improve the accuracy of pose detection.
  • the following describes how to determine the pose information of the target object in each frame of the first image in multiple frames of the first image based on grid data (mesh data).
  • the point cloud registration process is performed on the preset point cloud, the grid data corresponding to the preset point cloud is determined, and the relationship between the grid data and the preset point cloud The pose transformation information of .
  • the point cloud registration method can be based on the point cloud registration method, such as Nonrigid Iteration Closest Points.
  • the point cloud registration method is used to unify the preset point cloud whole_scan without regular topology (as shown in the right figure of Figure 5) into a regular topology
  • the grid data mesh (smooth plane representation as shown in the left figure of Figure 2).
  • non-rigid registration can be performed on the grid data mesh in the initial state and the preset point cloud whole_scan, and the mapping between the grid data mesh in the initial state and the preset point cloud whole_scan can be found, and the grid data in the initial state mesh Perform deformation to fit the preset point cloud whole_scan while maintaining the topology of the grid data mesh itself.
  • the pose transformation information trans_mesh of the preset point cloud whole_scan and the grid data mesh can be solved.
  • trans_mesh scale_1*rotation_1*mesh+translation_1.
  • the difference between the pose transformation information trans_mesh of the network data mesh and the preset point cloud whole_scan is the smallest (the pose transformation information trans_mesh of the network data mesh is infinitely close to the preset point cloud whole_scan, that is, the preset point cloud whole_scan and the grid When the data mesh is aligned, as shown in the left figure of Figure 7), record the pose transformation information trans_mesh of the grid data mesh and the preset point cloud whole_scan.
  • the relative relationship between the preset point cloud and the first point cloud can also be obtained synchronously, and the relative relationship includes the preset Pose transformation information between the point cloud and the first point cloud.
  • the method for determining the relative relationship between the preset point cloud and the first point cloud can be referred to above, and will not be repeated here.
  • the target object in the first image of each frame can be determined in step SB3 pose information.
  • step SB3 may include:
  • step SB31 according to the pose transformation information between the grid data and the preset point cloud, and the pose transformation information between the preset point cloud and the first point cloud, the pose transformation information between the grid data and the first point cloud;
  • step SB32 according to the internal parameters and external parameters of the camera array, the pose transformation information between the grid data and the first point cloud, determine the distance between the grid data and the first image of each frame Pose transformation information;
  • step SB33 according to the pose transformation information between the grid data and each frame of the first image, determine the pose information of the target object in each frame of the first image in the multiple frames of the first image.
  • FIG. 7 shows a schematic diagram of aligning grid data and a first point cloud according to an embodiment of the present disclosure.
  • Steps SB31-SB33 introduce the method of aligning the grid data and the first point cloud, as follows:
  • step SB31 according to the step grid data mesh and the pose transformation information trans_mesh of the preset point cloud whole_scan, and the pose transformation information trans_scan of the preset point cloud whole_scan and the first point cloud part_scan, the grid data mesh to the first point cloud is obtained.
  • the pose transformation information trans_part of the grid data mesh and the first point cloud part_scan is: scale_2*rotation_2*(scale_1*rotation_1*mesh+translation_1)+translation_2, which can align the grid data mesh with the first point cloud part_scan, As shown in the right figure of Figure 7.
  • step SB32 according to the calibration parameter trans_camera of the camera array obtained in step SB1, including the internal parameters and external parameters of the camera array, and the pose transformation information trans_part of the grid data mesh and the first point cloud part_scan, the grid data can be determined The pose transformation information trans_end of the mesh relative to the first image of the M frame;
  • step SB33 according to the pose transformation information trans_end of the grid data mesh relative to the M frames of the first image, the grid data mesh can be projected to the first images respectively, and the target object corresponding to the first image of each frame can be determined
  • the pose information, the pose information may include scaling information, rotation information and translation information.
  • FIG. 8 shows a schematic diagram of a projection effect according to an embodiment of the present disclosure. As shown in FIG. 8 , the grid data mesh can be rendered to each frame of the back image of the human head, and the pose information of the back image of the human head can be accurately obtained according to the embodiments of the present disclosure.
  • both the preset point cloud whole_scan and the first point cloud part_scan include a large number of points
  • direct analysis of these two point clouds requires large-scale calculations and consumes a large amount of computing resources.
  • key points can be marked on the preset point cloud whole_scan and the first point cloud part_scan respectively, and the alignment of the point cloud can be realized according to the alignment of the key points, and then the target in the first image of each frame can be determined based on the key points The pose information of the object.
  • the aligning the preset point cloud with the first point cloud and determining the relative relationship between the preset point cloud and the first point cloud may be Including: in the face area of the preset point cloud, generating annotation information including N preset key points; in the face area of the first point cloud, generating annotation information including K second key points information, K is an integer not greater than N; align the N key points of the preset point cloud with the K second key points of the first point cloud, and determine the preset point cloud and the first point cloud Relative relationship between point clouds.
  • N preset key points can be marked in the face area
  • Figure 5 shows the preset key points In the case of 10, it can correspond to the left/right outer corner of the eye, left/right inner corner of the eye, tip of the nose, left/right mouth corner, center of the chin, and left/right ear base.
  • K second key points can be marked in the face area.
  • Figure 6 shows the situation of preset key points. Since Figure 6 In the reconstructed first point cloud part_scan, part of the face area is missing, and the key points of the missing area can be ignored (for example, the four key points of the right outer corner of the eye, the inner corner of the right eye, the right mouth corner, and the right ear root point can be ignored.
  • N preset key points and/or K second key points are obtained in the face area, which can be marked through the received user labeling operation, or feature point detection (for example, based on deep learning methods, etc.) can be performed on the face. ) to obtain the marked points, and the present disclosure does not limit the method of generating marked information.
  • Fig. 5 and Fig. 6 are only for illustration, and the annotation information of the second key point can be generated at the position corresponding to the preset key point, and the key points at the corresponding position have the same point sequence.
  • the number of points and second key points, as well as the specific labeling positions of preset key points and second key points are not limited.
  • the preset key points corresponding to the K second keys can be selected from the N preset key points of the preset point cloud, and then according to the selected K preset key points Perform alignment processing with the K second key points to determine the relative relationship between the preset point cloud and the first point cloud.
  • alignment processing please refer to the alignment method above, and will not repeat it here.
  • steps SB1-SB3 for the scene of multiple frames of rear head images captured by a multi-camera array in a laboratory environment, the pose information of the target object in the first image of each frame can be accurately determined.
  • the embodiments of the present disclosure for some target object back images that cannot accurately provide clear semantic parts, it is also possible to accurately determine the three-dimensional model (preset point cloud) of the target object in the image relative to the target object, not only accurately Realize the pose of a single-frame image taken with the back to the camera (steps SA1-SA3), and also accurately realize the pose of multiple-frame images taken with the back to the camera array device (steps SB1-SB3), effectively solving the all-round problem of the target object Pose detection, regardless of the angle between the face orientation of the target object and the shooting direction, can achieve accurate pose detection.
  • the three-dimensional model preset point cloud
  • the present disclosure also provides pose detection devices, electronic equipment, computer-readable storage media, and programs, all of which can be used to implement any pose detection method provided in the present disclosure, and refer to the corresponding technical solutions and descriptions in the method section Corresponding records are not repeated here.
  • Fig. 9 shows a block diagram of a pose detection device according to an embodiment of the present disclosure. As shown in Fig. 9, the device includes:
  • An acquisition module 91 configured to acquire at least one frame of the first image, wherein the at least one frame of the first image includes at least one frame of the two-dimensional image of the back of the target object collected by the image acquisition device;
  • the determining module 92 is configured to determine the pose information of the target object in each of the at least one frame of the first image according to the at least one frame of the first image and a preset point cloud, wherein the preset The point cloud is a three-dimensional point cloud of the front of the target object.
  • the preset point cloud includes N preset key points, where N is an integer greater than 1, and when the image acquisition device is a single camera, the at least one frame first The image includes a single frame of the first image captured by the single camera; the determination module 92 is configured to: generate label information including N first key points in the target object area of the single frame of the first image; The preset point cloud is projected onto the single frame of the first image, and N first projection points corresponding to the N first key points are determined; the N first projection points and the N preset key points are simulated combined processing to obtain the pose information of the target object in the single frame first image.
  • the at least one frame of the first image includes: multiple frames of the first image and multiple frames of the second image, and the multiple frames of the second image are The multi-frame two-dimensional images of the front of the target object;
  • the determination module 92 is configured to: perform three-dimensional reconstruction on the multi-frame second images to obtain the preset point cloud, and carry out the multi-frame first image Three-dimensional reconstruction to obtain a first point cloud; according to the relative relationship between the preset point cloud and the first point cloud, determine the pose information of the target object in each frame of the first image among the multiple frames of first images.
  • the acquisition module 91 is configured to, when the image acquisition device is a camera array, before acquiring at least one frame of the first image, according to the checkerboard image acquired by the camera array , calibrate the internal parameters and external parameters of the camera array, the internal parameters include at least one of focal length or pixels, and the external parameters include at least one of camera position or camera rotation angle; the determination module 92 is used for: Performing three-dimensional reconstruction on the multiple frames of second images according to the internal parameters and the external parameters to obtain the preset point cloud; and, according to the internal parameters and the external parameters, performing three-dimensional reconstruction on the multiple frames of the second images performing 3D reconstruction on an image to obtain the first point cloud.
  • the relative relationship between the preset point cloud and the first point cloud includes pose transformation information between the preset point cloud and the first point cloud
  • the relative relationship between the preset point cloud and the first point cloud, and determining the pose information of the target object in each frame of the first image in the multiple frames of the first image include: according to the internal parameters and external parameters of the camera array , and the pose transformation information between the preset point cloud and the first point cloud, determining the pose transformation between the preset point cloud and each frame of the first image in the multiple frames of first images Information; according to the pose transformation information between the preset point cloud and each frame of the first image in the multiple frames of the first image, determine the target object in each frame of the first image in the multiple frames of the first image pose information.
  • the relative relationship between the preset point cloud and the first point cloud includes pose transformation information between the preset point cloud and the first point cloud
  • the determining Module 92 is also used to: perform point cloud registration processing on the preset point cloud, determine the grid data corresponding to the preset point cloud, and the pose between the grid data and the preset point cloud Transformation information: according to the relative relationship between the preset point cloud and the first point cloud, determining the pose information of the target object in each frame of the first image in the multiple frames of the first image includes: according to the The pose transformation information between the grid data and the preset point cloud, and the pose transformation information between the preset point cloud and the first point cloud, determine the grid data and the first point cloud Pose transformation information between point clouds; according to the internal parameters and external parameters of the camera array, the pose transformation information between the grid data and the first point cloud, determine the grid data and the Pose transformation information between each frame of the first image in the multiple frames of the first image; according to the pose transformation information between the grid data and each frame of the first image
  • the determining module 92 is further configured to: perform 3D reconstruction on the multiple frames of second images to obtain the preset point cloud, and perform three-dimensional reconstruction on the multiple frames of the first images After three-dimensional reconstruction is performed to obtain the first point cloud, an alignment process is performed on the preset point cloud and the first point cloud to determine a relative relationship between the preset point cloud and the first point cloud.
  • the aligning the preset point cloud with the first point cloud, and determining the relative relationship between the preset point cloud and the first point cloud include : In the face area of the preset point cloud, generate labeling information including N preset key points; in the face area of the first point cloud, generate labeling information including K second key points , K is an integer not greater than N; align the N key points of the preset point cloud with the K second key points of the first point cloud, and determine the preset point cloud and the first point Relative relationship between clouds.
  • the target object includes body parts of the human body; when the body parts of the human body include a human face, the preset key points include key points of human face organs; When the body parts include limbs, the preset key points include limb key points.
  • the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the methods described in the method embodiments above, and its specific implementation can refer to the description of the method embodiments above. For brevity, here No longer.
  • Embodiments of the present disclosure also provide a computer-readable storage medium, on which computer program instructions are stored, and the above-mentioned method is implemented when the computer program instructions are executed by a processor.
  • Computer readable storage media may be volatile or nonvolatile computer readable storage media.
  • An embodiment of the present disclosure also proposes an electronic device, including: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to invoke the instructions stored in the memory to execute the above method.
  • An embodiment of the present disclosure also provides a computer program product, including computer-readable codes, or a non-volatile computer-readable storage medium carrying computer-readable codes, when the computer-readable codes are stored in a processor of an electronic device When running in the electronic device, the processor in the electronic device executes the above method.
  • Electronic devices may be provided as terminals, servers, or other forms of devices.
  • FIG. 10 shows a block diagram of an electronic device 800 according to an embodiment of the present disclosure.
  • the electronic device 800 may be a terminal such as a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, or a personal digital assistant.
  • electronic device 800 may include one or more of the following components: processing component 802, memory 804, power supply component 806, multimedia component 808, audio component 810, input/output (I/O) interface 812, sensor component 814 , and the communication component 816.
  • the processing component 802 generally controls the overall operations of the electronic device 800, such as those associated with display, telephone calls, data communications, camera operations, and recording operations.
  • the processing component 802 may include one or more processors 820 to execute instructions to complete all or part of the steps of the above method. Additionally, processing component 802 may include one or more modules that facilitate interaction between processing component 802 and other components. For example, processing component 802 may include a multimedia module to facilitate interaction between multimedia component 808 and processing component 802 .
  • the memory 804 is configured to store various types of data to support operations at the electronic device 800 . Examples of such data include instructions for any application or method operating on the electronic device 800, contact data, phonebook data, messages, pictures, videos, and the like.
  • the memory 804 can be implemented by any type of volatile or non-volatile storage device or their combination, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), Magnetic Memory, Flash Memory, Magnetic or Optical Disk.
  • SRAM static random access memory
  • EEPROM electrically erasable programmable read-only memory
  • EPROM erasable Programmable Read Only Memory
  • PROM Programmable Read Only Memory
  • ROM Read Only Memory
  • Magnetic Memory Flash Memory
  • Magnetic or Optical Disk Magnetic Disk
  • the power supply component 806 provides power to various components of the electronic device 800 .
  • Power components 806 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for electronic device 800 .
  • the multimedia component 808 includes a screen providing an output interface between the electronic device 800 and the user.
  • the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user.
  • the touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensor may not only sense a boundary of a touch or swipe action, but also detect duration and pressure associated with the touch or swipe action.
  • the multimedia component 808 includes a front camera and/or a rear camera. When the electronic device 800 is in an operation mode, such as a shooting mode or a video mode, the front camera and/or the rear camera can receive external multimedia data. Each front camera and rear camera can be a fixed optical lens system or have focal length and optical zoom capability.
  • the audio component 810 is configured to output and/or input audio signals.
  • the audio component 810 includes a microphone (MIC), which is configured to receive external audio signals when the electronic device 800 is in operation modes, such as call mode, recording mode and voice recognition mode. Received audio signals may be further stored in memory 804 or sent via communication component 816 .
  • the audio component 810 also includes a speaker for outputting audio signals.
  • the I/O interface 812 provides an interface between the processing component 802 and a peripheral interface module, which may be a keyboard, a click wheel, a button, and the like. These buttons may include, but are not limited to: a home button, volume buttons, start button, and lock button.
  • Sensor assembly 814 includes one or more sensors for providing status assessments of various aspects of electronic device 800 .
  • the sensor component 814 can detect the open/closed state of the electronic device 800, the relative positioning of components, such as the display and the keypad of the electronic device 800, the sensor component 814 can also detect the electronic device 800 or a Changes in position of components, presence or absence of user contact with electronic device 800 , electronic device 800 orientation or acceleration/deceleration and temperature changes in electronic device 800 .
  • Sensor assembly 814 may include a proximity sensor configured to detect the presence of nearby objects in the absence of any physical contact.
  • Sensor assembly 814 may also include an optical sensor, such as a complementary metal-oxide-semiconductor (CMOS) or charge-coupled device (CCD) image sensor, for use in imaging applications.
  • CMOS complementary metal-oxide-semiconductor
  • CCD charge-coupled device
  • the sensor component 814 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor or a temperature sensor.
  • the communication component 816 is configured to facilitate wired or wireless communication between the electronic device 800 and other devices.
  • the electronic device 800 can access wireless networks based on communication standards, such as wireless networks (Wi-Fi), second-generation mobile communication technologies (2G), third-generation mobile communication technologies (3G), fourth-generation mobile communication technologies (4G ), the long-term evolution (LTE) of the universal mobile communication technology, the fifth generation mobile communication technology (5G) or their combination.
  • the communication component 816 receives broadcast signals or broadcast related information from an external broadcast management system via a broadcast channel.
  • the communication component 816 also includes a near field communication (NFC) module to facilitate short-range communication.
  • the NFC module may be implemented based on Radio Frequency Identification (RFID) technology, Infrared Data Association (IrDA) technology, Ultra Wideband (UWB) technology, Bluetooth (BT) technology and other technologies.
  • RFID Radio Frequency Identification
  • IrDA Infrared Data Association
  • UWB Ultra Wideband
  • Bluetooth Bluetooth
  • electronic device 800 may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable A programmable gate array (FPGA), controller, microcontroller, microprocessor or other electronic component implementation for performing the methods described above.
  • ASICs application specific integrated circuits
  • DSPs digital signal processors
  • DSPDs digital signal processing devices
  • PLDs programmable logic devices
  • FPGA field programmable A programmable gate array
  • controller microcontroller, microprocessor or other electronic component implementation for performing the methods described above.
  • a non-volatile computer-readable storage medium such as the memory 804 including computer program instructions, which can be executed by the processor 820 of the electronic device 800 to implement the above method.
  • This disclosure relates to the field of augmented reality.
  • acquiring the image information of the target object in the real environment and then using various visual correlation algorithms to detect or identify the relevant features, states and attributes of the target object, and thus obtain the image information that matches the specific application.
  • AR effect combining virtual and reality.
  • the target object may involve faces, limbs, gestures, actions, etc. related to the human body, or markers and markers related to objects, or sand tables, display areas or display items related to venues or places.
  • Vision-related algorithms can involve visual positioning, SLAM, 3D reconstruction, image registration, background segmentation, object key point extraction and tracking, object pose or depth detection, etc.
  • Specific applications can not only involve interactive scenes such as guided tours, navigation, explanations, reconstructions, virtual effect overlays and display related to real scenes or objects, but also special effects processing related to people, such as makeup beautification, body beautification, special effect display, virtual Interactive scenarios such as model display.
  • the relevant features, states and attributes of the target object can be detected or identified through the convolutional neural network.
  • the above-mentioned convolutional neural network is a network model obtained by performing model training based on a deep learning framework.
  • FIG. 11 shows a block diagram of an electronic device 1900 according to an embodiment of the present disclosure.
  • electronic device 1900 may be provided as a server.
  • electronic device 1900 includes processing component 1922 , which further includes one or more processors, and a memory resource represented by memory 1932 for storing instructions executable by processing component 1922 , such as application programs.
  • the application programs stored in memory 1932 may include one or more modules each corresponding to a set of instructions.
  • the processing component 1922 is configured to execute instructions to perform the above method.
  • Electronic device 1900 may also include a power supply component 1926 configured to perform power management of electronic device 1900, a wired or wireless network interface 1950 configured to connect electronic device 1900 to a network, and an input-output (I/O) interface 1958 .
  • the electronic device 1900 can operate based on the operating system stored in the memory 1932, such as the Microsoft server operating system (Windows ServerTM), the graphical user interface-based operating system (Mac OS XTM) introduced by Apple Inc., and the multi-user and multi-process computer operating system (UnixTM). ), a free and open source Unix-like operating system (LinuxTM), an open source Unix-like operating system (FreeBSDTM), or similar.
  • a non-transitory computer-readable storage medium such as the memory 1932 including computer program instructions, which can be executed by the processing component 1922 of the electronic device 1900 to implement the above method.
  • the present disclosure can be a system, method and/or computer program product.
  • a computer program product may include a computer readable storage medium having computer readable program instructions thereon for causing a processor to implement various aspects of the present disclosure.
  • a computer readable storage medium may be a tangible device that can retain and store instructions for use by an instruction execution device.
  • a computer readable storage medium may be, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
  • Computer-readable storage media include: portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or flash memory), static random access memory (SRAM), compact disc read only memory (CD-ROM), digital versatile disc (DVD), memory stick, floppy disk, mechanically encoded device, such as a printer with instructions stored thereon A hole card or a raised structure in a groove, and any suitable combination of the above.
  • RAM random access memory
  • ROM read-only memory
  • EPROM erasable programmable read-only memory
  • flash memory static random access memory
  • SRAM static random access memory
  • CD-ROM compact disc read only memory
  • DVD digital versatile disc
  • memory stick floppy disk
  • mechanically encoded device such as a printer with instructions stored thereon
  • a hole card or a raised structure in a groove and any suitable combination of the above.
  • computer-readable storage media are not to be construed as transient signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., pulses of light through fiber optic cables), or transmitted electrical signals.
  • Computer-readable program instructions described herein may be downloaded from a computer-readable storage medium to a respective computing/processing device, or downloaded to an external computer or external storage device over a network, such as the Internet, a local area network, a wide area network, and/or a wireless network.
  • the network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and/or edge servers.
  • a network adapter card or a network interface in each computing/processing device receives computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing/processing device .
  • Computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or Source or object code written in any combination, including object-oriented programming languages—such as Smalltalk, C++, etc., and conventional procedural programming languages—such as the “C” language or similar programming languages.
  • Computer readable program instructions may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server implement.
  • the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (such as via the Internet using an Internet service provider). connect).
  • LAN local area network
  • WAN wide area network
  • an electronic circuit such as a programmable logic circuit, field programmable gate array (FPGA), or programmable logic array (PLA)
  • FPGA field programmable gate array
  • PDA programmable logic array
  • These computer-readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine such that when executed by the processor of the computer or other programmable data processing apparatus , producing an apparatus for realizing the functions/actions specified in one or more blocks in the flowchart and/or block diagram.
  • These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause computers, programmable data processing devices and/or other devices to work in a specific way, so that the computer-readable medium storing instructions includes An article of manufacture comprising instructions for implementing various aspects of the functions/acts specified in one or more blocks in flowcharts and/or block diagrams.
  • each block in a flowchart or block diagram may represent a module, a portion of a program segment, or an instruction that includes one or more Executable instructions.
  • the functions noted in the block may occur out of the order noted in the figures. For example, two blocks in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations can be implemented by a dedicated hardware-based system that performs the specified function or action , or may be implemented by a combination of dedicated hardware and computer instructions.
  • the computer program product can be specifically realized by means of hardware, software or a combination thereof.
  • the computer program product is embodied as a computer storage medium, and in another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (Software Development Kit, SDK) etc. wait.
  • a software development kit Software Development Kit, SDK

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Geometry (AREA)
  • Computer Graphics (AREA)
  • Image Analysis (AREA)

Abstract

本公开涉及一种位姿检测方法及装置、电子设备和存储介质，所述方法包括：获取至少一帧第一图像，其中，该至少一帧第一图像的每个包括图像采集设备采集的目标对象背面的至少一帧二维图像，根据至少一帧第一图像，以及预设的目标对象正面的三维点云，确定所述至少一帧第一图像的每个中目标对象的位姿信息。

Description

位姿检测方法及装置、电子设备和存储介质
相关申请的交叉引用
本公开要求于2022年02月17日提交的、申请号为202210147653.7的中国专利申请的优先权，该申请以引用的方式并入本文中。
技术领域
本公开涉及计算机技术领域，尤其涉及一种位姿检测方法及装置、电子设备和存储介质。
背景技术
基于图像的目标三维重建已经成为十分流行的研究方向，并且在目标的检测与识别、目标驱动、增强现实等方向得到广泛应用，与之对应的，求解目标对象的三维模型和二维图像间的相对位姿(Pose)，也形成了一套比较成熟的逻辑方案。
但是，在对目标对象进行三维重建的过程中，当目标对象背对镜头时，由于目标对象背面没有明确的语义部位，例如，以人头三维重建为例，当人头后脑勺的区域朝向镜头时，后脑勺上没有例如眼角、嘴角、鼻尖等拥有明确语义的部位，所以，在进行目标对象的三维模型(人头三维模型)相对于背面二维图像(例如后脑勺图像)的位姿检测时会遇到较大的困难。
发明内容
本公开提出了一种位姿检测技术方案。
根据本公开的一方面，提供了一种位姿检测方法，包括：获取至少一帧第一图像，其中，所述至少一帧第一图像包括图像采集设备采集的目标对象背面的至少一帧二维图像；根据所述至少一帧第一图像，以及预设点云，确定所述至少一帧第一图像的每个中所述目标对象的位姿信息，其中，所述预设点云为所述目标对象正面的三维点云。
在一种可能的实现方式中，所述预设点云包括N个预设关键点，N为大于1的整数，在图像采集设备为单相机的情况下，所述至少一帧第一图像包括所述单相机拍摄的单帧第一图像；所述根据所述至少一帧第一图像，以及预设点云，确定所述至少一帧第一图像中所述目标对象的位姿信息，包括：在所述单帧第一图像的目标对象区域，生成包含N个第一关键点的标注信息；将所述预设点云投影至所述单帧第一图像，确定N个第一关键点对应的N个第一投影点；对所述N个第一投影点与所述N个预设关键点进行拟合处理，得到所述单帧第一图像中目标对象的位姿信息。
在一种可能的实现方式中，在图像采集设备为相机阵列的情况下，所述至少一帧第一图像包括：多帧第一图像以及多帧第二图像，所述多帧第二图像为目标对象正面的多帧二维图像；所述根据所述至少一帧第一图像，以及预设点云，确定所述至少一帧第一图像的每个中所述目标对象的位姿信息，包括：对所述多帧第二图像进行三维重建，得到所述预设点云，以及对所述多帧第一图像进行三维重建，得到第一点云；根据所述预设点云和所述第一点云的相对关系，确定所述多帧第一图像中每帧第一图像中目标对象的位姿信息。
在一种可能的实现方式中，在图像采集设备为相机阵列的情况下，在所述获取至少一帧第一图像之前，包括：根据所述相机阵列采集的棋盘格图像，标定所述相机阵列的内参数和外参数，所述内参数包括焦距或像素中的至少一个，所述外参数包括相机位置或相机旋转角度中的至少一个；所述对所述多帧第二图像进行三维重建，得到所述预设 点云，包括：根据所述内参数和所述外参数，对所述多帧第二图像进行三维重建，得到所述预设点云；以及，所述对所述多帧第一图像进行三维重建，得到第一点云，包括：根据所述内参数和所述外参数，对所述多帧第一图像进行三维重建，得到所述第一点云。
在一种可能的实现方式中，所述预设点云和所述第一点云的相对关系包括所述预设点云与所述第一点云之间的位姿变换信息，所述根据所述预设点云和所述第一点云的相对关系，确定所述多帧第一图像的每帧第一图像中所述目标对象的位姿信息，包括：根据所述相机阵列的内参数和外参数、以及所述预设点云与所述第一点云之间的位姿变换信息，确定所述预设点云与所述多帧第一图像中每帧第一图像之间的位姿变换信息；根据所述预设点云与所述多帧第一图像中每帧第一图像之间的位姿变换信息，确定所述多帧第一图像中每帧第一图像中所述目标对象的位姿信息。
在一种可能的实现方式中，所述预设点云和所述第一点云的相对关系包括所述预设点云与所述第一点云之间的位姿变换信息，所述方法还包括：对所述预设点云进行点云注册处理，确定所述预设点云对应的网格数据，以及所述网格数据与所述预设点云之间的位姿变换信息；所述根据所述预设点云和所述第一点云的相对关系，确定所述多帧第一图像中每帧第一图像中目标对象的位姿信息，包括：根据所述网格数据与所述预设点云之间的位姿变换信息，以及所述预设点云与所述第一点云之间的位姿变换信息，确定所述网格数据与所述第一点云之间的位姿变换信息；根据所述相机阵列的内参数和外参数、所述网格数据与所述第一点云之间的位姿变换信息，确定所述网格数据与所述多帧第一图像中每帧第一图像之间的位姿变换信息；根据所述网格数据与所述多帧第一图像中每帧第一图像之间的位姿变换信息，确定所述多帧第一图像中每帧第一图像中目标对象的位姿信息。
在一种可能的实现方式中，所述对所述多帧第二图像进行三维重建，得到所述预设点云，以及对所述多帧第一图像进行三维重建，得到第一点云之后，所述方法还包括：对所述预设点云与所述第一点云进行对齐处理，确定所述预设点云与所述第一点云之间的相对关系。
在一种可能的实现方式中，所述对所述预设点云与所述第一点云进行对齐处理，确定所述预设点云与所述第一点云之间的相对关系，包括：在所述预设点云的人脸区域中，生成包含N个预设关键点的标注信息；在所述第一点云的人脸区域中，生成包含K个第二关键点的标注信息，K为不大于N的整数；对所述预设点云的N个关键点与所述第一点云的K个第二关键点进行对齐处理，确定预设点云与所述第一点云之间的相对关系。
在一种可能的实现方式中，所述目标对象包括人体的身体部位；所述人体的身体部位包括人脸的情况下，所述预设关键点包括人脸器官的关键点；所述人体的身体部位包括肢体的情况下，所述预设关键点包括肢体关键点。
根据本公开的一方面，提供了一种位姿检测装置，包括：获取模块，用于获取至少一帧第一图像，所述第一图像包括图像采集设备采集的目标对象背面的二维图像；确定模块，用于根据所述至少一帧第一图像，以及预设点云，确定所述第一图像中目标对象的位姿信息，其中，所述预设点云为目标对象正面的三维点云。
根据本公开的一方面，提供了一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为调用所述存储器存储的指令，以执行上述方法。
根据本公开的一方面，提供了一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述方法。
在本公开实施例中，能够获取待处理的至少一帧第一图像，该第一图像包括图像采集设备采集的目标对象背面的二维图像，根据至少一帧第一图像，以及预设的目标对象正面的三维点云，确定所述第一图像中目标对象的位姿信息，通过这种方式，不仅能准 确实现背对相机拍摄的单帧图像的位姿，还能准确实现背对相机阵列设备拍摄的多帧图像的位姿，有效解决目标对象全方位的位姿检测，不管目标对象的面部朝向与拍摄方向的角度呈现什么角度，均能实现准确地位姿检测。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，而非限制本公开。根据下面参考附图对示例性实施例的详细说明，本公开的其它特征及方面将变得清楚。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，这些附图示出了符合本公开的实施例，并与说明书一起用于说明本公开的技术方案。
图1示出根据本公开实施例的位姿检测方法的流程图。
图2示出根据本公开实施例的预设关键点和第一关键点的示意图。
图3示出根据本公开实施例的投影效果的示意图。
图4示出根据本公开实施例的标定板的示意图。
图5示出根据本公开实施例的预设点云获取方式的示意图。
图6示出根据本公开实施例的第一点云获取方式的示意图。
图7示出根据本公开实施例的网格数据和第一点云对齐的示意图。
图8示出根据本公开实施例的投影效果的示意图。
图9示出根据本公开实施例的位姿检测装置的框图。
图10示出根据本公开一实施例的一种电子设备的框图。
图11示出根据本公开另一实施例的一种电子设备的框图。
具体实施方式
以下将参考附图详细说明本公开的各种示例性实施例、特征和方面。附图中相同的附图标记表示功能相同或相似的元件。尽管在附图中示出了实施例的各种方面，但是除非特别指出，不必按比例绘制附图。
在这里专用的词“示例性”意为“用作例子、实施例或说明性”。这里作为“示例性”所说明的任何实施例不必解释为优于或好于其它实施例。
本文中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本文中术语“至少一种”表示多种中的任意一种或多种中的至少两种的任意组合，例如，包括A、B、C中的至少一种，可以表示包括从A、B和C构成的集合中选择的任意一个或多个元素。
另外，为了更好地说明本公开，在下文的具体实施方式中给出了众多的具体细节。本领域技术人员应当理解，没有某些具体细节，本公开同样可以实施。在一些实例中，对于本领域技术人员熟知的方法、手段、元件和电路未作详细描述，以便于凸显本公开的主旨。
对目标对象进行位姿检测可广泛应用于各种场景中，其中，目标对象可以包括人体的身体部位等，例如，目标对象为正面拥有明确语义部位(例如眼角、嘴角、鼻尖等)而背面缺少明确语义部位的对象。应当理解，本公开实施例的目标对象不限于人体的身体部位(例如人头、肢体等)，任何正面拥有明确语义部位并且背面缺少明确语义部位的物体，可作为目标对象。为了便于说明，后续各公开实施例均以目标对象为人头为例进行说明，目标对象为其他对象的情况可以根据后续各公开实施例进行灵活扩展，本公开对此不作限制。
以人头的位姿检测为例，对于人脸/人头360度重建任务的场景，为了提高人脸/人头 360度重建的准确度，需要对360度范围内任意角度的人头图像进行位姿检测，确定人脸/人头相对于正面和背面图片的位姿，以便提供该360度重建任务需要的训练数据。或者，在一些需要以人头朝向作为重要判定依据的场景，也需要对人头图像中的人头进行位姿检测，比如，在安全驾驶的检测场景中，如果车载摄像头设置于驾驶位后方，需要基于采集的驾驶员的后脑勺图像判定驾驶员是否在左顾右盼等。或者，在基于人脸位姿进行增强现实(Augmented Reality，AR)交互的应用场景，需要对拍摄的人脸图像进行位姿检测，比如，对于应用人脸贴纸特效的场景，用户想加入发卡等装饰品特效时，需要准确定位任意角度的人头，减少在人头背对拍摄镜头时出现贴纸特效丢失的情况。其中，基于人脸位姿进行增强现实AR交互的应用场景，除了人脸贴纸特效，还可以包括虚拟妆容处理、人脸转换动物脸特效等，本公开对此不作限制。
对待处理图像中的目标对象进行位姿检测，可估计目标对象相对于某个参考基准的朝向或姿态，其中，参考基准可以是预设的目标对象的三维模型，在实际使用中，可以基于实际需求选取目标对象位姿的参考基准。
然而，相关技术中，对待处理图像中的目标对象进行位姿检测，主要集中于目标对象正面面对相机，左右旋转90度之内的位姿检测，即目标对象的面部朝向与拍摄方向(面部与图像采集设备之间的方向)的角度在0～90度，以及270度～360度之间。例如，在对待处理图像中的目标对象进行位姿检测时，可以首先筛选出目标对象的面部朝向与拍摄方向的角度在0～90度，以及270度～360度之间的各角度图像，由于筛选出的目标对象的图像具有明确语义部位(例如眼角、嘴角、鼻尖等)，可以分别从筛选出的目标对象的图像中提取出关键点，基于关键点拟合得到目标对象的三维模型，然后可以以该三维模型作为参考基准，确定图像中的目标对象相对于该三维模型的位姿。
其中，三维模型可以是一个标准化的模型，例如三维密集点云，该标准化的模型可以具有拓扑结构，即模型中各点的连接关系是确定的，且模型中各点与目标对象的各个部位也是绑定的。
上述位姿检测方法需要提取图像中目标对象的关键点，基于关键点拟合目标对象的三维模型，进而才能确定图像中的目标对象相对于三维模型的位姿。
如果目标对象的面部朝向与拍摄方向的角度在0～90度，以及270度～360度之间，拍摄到的是旋转角度较小的目标对象(例如，正脸图像)，图像中的目标对象具有明确语义部位(例如眼角、嘴角、鼻尖等)，关键点较多的情况，可以准确拟合三维模型并估计目标对象相对于三维模型的位姿。
但是，如果目标对象的面部朝向与拍摄方向的角度在90度～270度之间，拍摄到的是旋转角度较大的目标对象(例如，后脑勺图像)，图像中的目标对象区域仅包括很少的明确语义部位(例如眼角、嘴角、鼻尖等)，甚至有些目标对象图像完全没有包括明确语义部位，可提取到的关键点数量很少甚至没有，在此情况下，会无法准确拟合得到三维模型，导致得到的目标对象相对于三维模型的位姿检测结果误差很大。可见，相关技术不能有效解决目标对象在任意拍摄角度下的准确位姿检测，尤其不能有效解决在目标对象的背面处于拍摄角度下的准确位姿检测。
应当理解，0～90度，90度～270度，270度～360度仅作为一个参考范围，本公开的实施例并不限于以上给出的数值范围。
有鉴于此，本公开提供了一种位姿检测技术方案，能够获取待处理的至少一帧第一图像，该第一图像包括图像采集设备采集的目标对象背面的二维图像，根据至少一帧第一图像，以及预设的目标对象正面的三维点云，确定所述第一图像中目标对象的位姿信息，通过这种方式，不仅能准确实现背对相机拍摄的单帧图像的位姿，还能准确实现背对相机阵列设备拍摄的多帧图像的位姿，有效解决目标对象全方位的位姿检测，不管目标对象的面部朝向与拍摄方向的角度呈现什么角度，均能实现准确地位姿检测。
图1示出根据本公开实施例的位姿检测方法的流程图，该位姿检测方法可以由终端 设备或服务器等电子设备执行，终端设备可以为用户设备(User Equipment，UE)、移动设备、用户终端、蜂窝电话、无绳电话、个人数字助理(Personal Digital Assistant，PDA)、手持设备、计算设备、车载设备、可穿戴设备等，所述方法可以通过处理器调用存储器中存储的计算机可读指令的方式来实现。或者，可通过服务器执行所述方法。
下面以电子设备作为执行主体为例对本公开实施例的位姿检测方法进行说明。
如图1所示，所述位姿检测方法包括：
在步骤S1中，获取至少一帧第一图像，其中，所述至少一帧第一图像包括图像采集设备采集的目标对象背面的至少一帧二维图像；
在一种可能的实现方式中，目标对象，可以包括正面(例如包括面部)拥有明确语义部位(例如包括眼角、嘴角、鼻尖等)，而背面(例如包括后脑勺)缺少明确语义部位的对象。应当理解，本公开实施例的目标对象不限于头部，任何正面拥有明确语义部位并且背面缺少明确语义部位的物体，可作为目标对象。后续各公开实施例均以目标对象为人头为例进行说明，目标对象为其他对象的情况可以根据后续各公开实施例进行灵活扩展，本公开对此不作限制。
在一种可能的实现方式中，待处理的第一图像可以是预先获取的包括目标对象的图像，或者是利用图像采集设备对目标对象进行拍摄时获取的图像。例如，在电子设备为手机、电脑等终端设备的场景下，可以由终端设备连接的图像采集设备采集包括目标对象的第一图像，或者从终端设备的相册中选择包括了目标对象的第一图像、或者从终端设备中安装的其他应用程序中接收包括目标对象的第一图像。
在一种可能的实现方式中，图像采集设备，可以是单相机，也可以是由处于不同空间位置的多相机构成的相机阵列。
基于不同的图像采集设备，一次拍摄可以获取一帧或多帧第一图像。如果图像采集设备为单相机，在单相机与目标对象的相对位置和角度无变化的情况下，可获取一帧待处理的第一图像，该第一图像为对应这一拍摄角度下目标对象的二维图像。
如果图像采集设备为相机阵列，例如针对由处于不同空间位置的M个(M为大于1的整数)相机构成的相机阵列，在相机阵列与目标对象的相对位置和角度无变化的情况下，可获取M帧待处理的第一图像，每帧第一图像分别对应各自相机的拍摄角度下的目标对象的二维图像。
其中，本公开实施例的图像采集设备除了可以是相机，也可以是摄像头、摄像机、扫描仪、手机、平板电脑等带有拍照功能的设备，本公开对此不作限制。
在一种可能的实现方式中，通过调整图像采集设备的角度或位置，图像采集设备可以采集目标对象在各角度下的图像。例如，在目标对象正对图像采集设备的情况下，也即目标对象的面部朝向与拍摄方向的角度在0～90度，以及270度～360度之间，可采集关于目标对象正面的待处理的至少一帧第一图像；或者，在目标对象背对图像采集设备的情况下，也即目标对象的面部朝向与拍摄方向的角度在90度～270度之间，可采集关于目标对象背面的待处理的至少一帧第一图像。
应当理解，对比相关技术中仅针对目标对象的正面图像进行准确的位姿检测，本公开实施例的位姿检测方法，不仅可以应用于目标对象正面的图像，还可以应用于目标对象背面的图像，对目标对象背面的二维图像进行位姿检测。在此情况下，本公开实施例侧重介绍在第一图像为目标对象背面的二维图像的情况下，对获取的待处理的至少一帧第一图像中目标对象进行位姿检测的方法。
在步骤S2中，根据所述至少一帧第一图像，以及预设点云，确定所述至少一帧第一图像的每个中所述目标对象的位姿信息。
其中，所述预设点云为目标对象正面的三维点云，所述位姿信息包括以下信息中的至少一种：缩放信息、旋转信息和平移信息。所述缩放信息可表示第一图像中目标对象与三维点云之间放大或缩小比例，所述旋转信息可表示第一图像中目标对象与三维点云 之间的旋转角度，所述平移信息可表示第一图像中目标对象与三维点云之间的移动的方向和距离。
应当理解，由于第一图像中的目标对象为二维数据，三维点云为三维数据，为了确定第一图像中目标对象的缩放信息、旋转信息和/或平移信息，可计算第一图像中目标对象与三维点云的二维投影之间的缩放信息、旋转信息和/或平移信息，本公开对具体二维数据与三维数据之间确定缩放信息、旋转信息和/或平移信息的具体方式不作限制。
在确定了第一图像中目标对象的位姿信息，可以将确定的目标对象的位姿信息应用在影视，游戏，虚拟社交等基于增强现实AR的互娱场景中，目标对象的位姿信息有利于判断现实中目标对象与虚拟信息(例如包括虚拟场景、虚拟物体等)的相对位置，实现目标对象与虚拟信息间的结合或交互，例如，在AR切水果游戏中，利用目标对象的位姿信息，可以实现多个虚拟水果以不同角度飞向目标对象。又例如，在AR换装应用中，可利用目标对象的位姿信息调整虚拟衣物的显示角度和位置，提高换装效果。例如，在AR特效生成应用中，目标对象的位姿信息有利于虚拟特效和目标对象的结合，比如虚拟妆容处理、人脸转换动物脸特效、人脸贴纸特效等。本公开对目标对象的位姿信息的应用场景不作限制。
在一种可能的实现方式中，预设点云为用于描述目标对象正面的三维点云，例如包括在世界坐标系(World Coordinate System)下构建的目标对象正面的三维点云，该预设点云可用于作为目标对象位姿评估参考基准。
其中，可以从电子设备内存中读取预先储存的预设点云，也可以对目标对象的多帧正面的二维图像进行三维重建，进而得到目标对象的预设点云，还可以对目标对象进行测绘得到预设点云，本公开对确定预设点云的获取方式不做限制。
在一种可能的实现方式中，可分析至少一帧第一图像中目标对象的图像坐标点与目标对象正面的三维点云(预设点云)的对应关系，确定至少一帧第一图像中目标对象的位姿信息，该位姿信息可包括第一图像中目标对象相对于预设点云的缩放信息、旋转信息和平移信息中的至少一种。
由于该第一图像中的目标对象处于二维空间，预设点云处于三维空间，为了便于分析第一图像中目标对象的图像坐标点与预设点云的对应关系，可以利用投影方式，将三维的预设点云投影至二维的图像空间，分析第一图像与投影后的预设点云的对应关系，进而确定至少一帧第一图像中目标对象的位姿信息；或者，还可以根据至少一帧第一图像，对至少一帧第一图像进行三维重建，以便重构目标对象的三维点云(即第一点云)，再分析重构后的第一点云与预设点云的对应关系，进而确定至少一帧第一图像中目标对象的位姿信息。
其中，为了提高位姿评估的效率，可在预设点云拥有明确语义部位处标注多个预设关键点，在至少一帧第一图像上或基于至少一帧第一图像重构的第一点云上标注与预设关键点具有相同点序的关键点。然后，可分析预设点云的关键点，与至少一帧第一图像上或基于至少一帧第一图像重构的第一点云上关键点的对应关系，进而确定至少一帧第一图像中目标对象的位姿信息。
例如，对于背对单相机采集的单帧第一图像，可通过分析预设点云的预设关键点与单帧第一图像的第一关键点的对应关系，对预设点云的预设关键点的第一投影点与单帧第一图像的第一关键点进行拟合处理，得到单帧第一图像中目标对象的位姿信息，所述位姿信息包括以下信息中的至少一种：缩放信息、旋转信息和平移信息。所述缩放信息可表示单帧第一图像中目标对象与预设点云的投影之间放大或缩小比例，所述旋转信息可表示单帧第一图像中目标对象与预设点云的投影之间的旋转角度，所述平移信息可表示单帧第一图像中目标对象与预设点云的投影之间的移动的方向和距离。
或者，对于背对相机阵列采集的多帧第一图像，可通过三维点云重建技术构建背面目标对象对应的第一点云，可在第一点云上标注第二关键点，可将预设关键点与第二关 键点对齐，也即将预设点云与第一点云对齐，以获取预设点云与第一点云之间的位姿变换信息，进而可根据该位姿变换信息，确定多帧第一图像中目标对象的位姿信息。所述位姿信息包括以下信息中的至少一种：缩放信息、旋转信息和平移信息。所述缩放信息可表示每帧第一图像中目标对象与预设点云的投影之间放大或缩小比例，所述旋转信息可表示每帧第一图像中目标对象与预设点云的投影之间的旋转角度，所述平移信息可表示每帧第一图像中目标对象与预设点云的投影之间的移动的方向和距离。
通过步骤S1～S2，针对一些无法准确提供明确语义部位的目标对象背面图像，也可以准确地确定出图像中目标对象相对于目标对象的三维模型(预设点云)的位姿信息，例如，通过步骤S1～S2，不仅能准确确定背对相机拍摄的单帧图像中目标对象的位姿，还能准确确定背对相机阵列设备拍摄的多帧图像中目标对象的位姿，有效实现目标对象全方位的位姿检测，不管目标对象的面部朝向与拍摄方向的角度呈现什么角度，均能实现准确地实现位姿检测。
下面以人头作为目标对象为例，针对背对相机(非相机阵列)拍摄的单帧背面人头图像的场景，以及多相机阵列拍摄的多帧背面人头图像的场景，分别对本公开实施例的位姿检测方法进行示例性的说明。
在一种可能的实现方式中，在步骤S1中，在图像采集设备为单相机的情况下，可获取待处理的至少一帧第一图像，所述至少一帧第一图像包括所述单相机拍摄的单帧第一图像。
其中，单帧第一图像可以是在自然(In The Wild)场景下，人头背对单相机时拍摄的二维图像，示例性的，该单帧第一图像可以是人脸朝向与拍摄方向的角度在90度～270度之间任一角度的图像。应当理解，90度～270度仅作为一个参考范围，本公开的实施例并不限于以上给出的数值范围。
其中，自然场景即在现实世界(例如工作、学习、生活)中的真实场景，例如可包括在公园、办公室、学校、商场等环境下拍摄的单帧第一图像，自然场景下拍摄的单帧第一图像的背景可以是多种多样的。
在步骤S1获取到单相机拍摄的单帧第一图像，可在步骤S2中，根据单帧第一图像的第一关键点，以及预设点云的预设关键点，确定第一图像中人头的位姿信息，步骤S2可包括：
在步骤SA1中，在单帧第一图像的目标对象区域，生成包含N个第一关键点的标注信息，所述标注信息可包括所述第一关键点的位置信息；
在步骤SA2中，将所述预设点云投影至所述单帧第一图像，确定N个第一关键点对应的N个第一投影点；
在步骤SA3中，对所述N个第一投影点与所述N个预设关键点进行拟合处理，得到所述单帧第一图像中目标对象的位姿信息。示例性的，图2示出根据本公开实施例的预设关键点和第一关键点的示意图。如图2所示，左图中的人头三维模型为预设点云(即人头正面的三维点云)，该预设点云可包括N个预设关键点，例如，在预设关键点数量为10个(N＝10)的情况下，预设关键点包括人脸器官的关键点，所述关键点包括：左/右外侧眼角点，左/右内侧眼角点，鼻尖点，左/右嘴角点，下巴中心点，左/右侧耳根点，其中，图2左图中数字标号顺序代表预设关键点的点序。
应当理解，上述仅以人头为示例，目标对象可包括人体的身体部位，所述人体的身体部位包括人脸的情况下，所述预设关键点包括人脸器官的关键点；所述人体的身体部位包括肢体的情况下，所述预设关键点包括肢体关键点，本公开对预设关键点不作具体限制。
如图2所示，右侧的人头背面图像为步骤S1中获取的单相机拍摄的单帧第一图像，可在步骤SA1中，可根据N个预设关键点，可在单帧第一图像的人头区域，生成包含N个第一关键点的标注信息，比如，可在单帧第一图像的人头区域根据接收到的用户操 作信息来标注N个第一关键点，生成包含N个第一关键点的标注信息；也可以利用预设的神经网络模型，将单帧第一图像输入该神经网络模型，生成包含N个第一关键点的标注信息，其中，预设的神经网络模型也即神经网络经过大量的样本训练(学习)后，得到的能够基于单帧第一图像的人头区域进行标注第一关键点的神经网络模型。本公开对标注信息的生成方法不作限制。
其中，N个第一关键点在人头背面(后脑勺)的标注位置，可对应于N个预设关键点的标注位置，并且与N个预设关键点具有相同的点序。例如，在第一关键点数量为10个(N＝10)的情况下，第一关键点可分别对应左/右外侧眼角点，左/右内侧眼角点，鼻尖点，左/右嘴角点，下巴中心点，左/右侧耳根点，图2右图中数字标号顺序代表第一关键点的点序。
应当理解，由于人头背面图像也可以看到耳朵部分，为了预估脸的大小，提高位姿检测的准确性，在选取待标注的第一关键点和预设关键点时，可至少包括左/右侧耳根点。其中，第一关键点与预设关键点的数量N的取值越大，本公开实施例的位姿检测方法的计算量越大，相应的位姿检测也会越准确。本公开对第一关键点的标注数量N的具体取值，以及标注的位置不作限制。
在确定了单帧第一图像中标注的N个第一关键点，以及预设点云的N个预设关键点，可在步骤SA2～SA3中，将如图2左图所示的三维预设点云投影至如图2右图所示的二维单帧第一图像，使预设点云上N个预设关键点的第一投影点与单帧第一图像中N个第一关键点相拟合，得到所述单帧第一图像对应目标对象的位姿信息。
例如，假设P3d代表预设点云中的N个预设关键点，即：(x 1，y 1，z 1)～(x N，y N，z N)，P2d代表单帧第一图像中的N个第一关键点，即：(x 1′，y 1′)～(x N′，y N′)。
基于三维空间的N个预设关键点(x 1，y 1，z 1)～(x N，y N，z N)和N个二维平面的第一关键点(x 1′，y 1′)～(x N′，y N′)，可以基于投影构建等式，使预设点云上N个预设关键点P3d的第一投影点与单帧第一图像中N个第一关键点P2d相拟合，即：
P2d＝Project(scale*rotation*P3d+translation)      (1)。
在公式(1)中，Project代表投影函数，可用于对目标对象进行降维处理，能够将三维空间的目标对象投影至二维平面，具体的投影方式可包括正交投影、透视投影等方式，本公开对投影方式的种类不做限制。
在公式(1)中，为了使预设点云上N个预设关键点P3d的第一投影点与单帧第一图像中N个第一关键点P2d相拟合，也即使投影后的N个预设关键点P3d与N个第一关键点P2d相重合或无限接近，在投影过程中，可以对预设点云的N个关键点p3d进行位姿变换，例如可包括缩放变换scale、旋转变换rotation和平移变换translation，并将使公式(1)中等式成立的位姿变换的取值，作为单帧第一图像中目标对象的位姿信息，该位姿信息可包括求解公式(1)得到的缩放信息scale、旋转信息rotation、以及平移信息translation。
示例性的，缩放变换scale可以是一个大于0的参数，旋转变换rotation可以是一个3*3的矩阵，平移变换可以是一个2*1的矩阵。具体实施中，有关旋转变换和平移变化的矩阵参数并不限制于上述示例说明。
其中，在将预设点云投影至所述单帧第一图像，确定N个第一关键点对应的N个第一投影点；可基于最小二乘法等拟合处理方法，使所述N个第一投影点与所述N个预设关键点相拟合，得到所述单帧第一图像中目标对象的位姿信息。
举例来说，可以将预设点云投影至所述单帧第一图像，确定N个第一关键点p3d：(x 1，y 1，z 1)～(x N，y N，z N)所对应的N个第一投影点为：Project(scale*rotation*P3d+translation)，也即：(px 1，py 1)～(px N，py N)；
为了提高计算效率，可利用线性优化的方法，例如最小二乘法(least sqaure method)， 对公式(1)进行求解，得到单帧第一图像中目标对象的位姿信息，即：
Figure PCTCN2022134852-appb-000001
在公式(2)，(x i′，y i′)代表第一关键点，(px i，py i)代表第一投影点，将使公式(2)取值最小的缩放信息scale、旋转信息rotation、以及平移信息translation，确定为单帧第一图像中目标对象的位姿信息。
通过这种方式，假设以目标对象是人头为例，对于单张后脑勺图像，可在后脑勺生成包含N个第一关键点的标注信息，使N个第一关键点和正面人脸模型(预设点云)的N个第一投影点具有相同的点序，然后直接根据成对的这N组点的拟合处理(例如利用最小二乘法)，求解单帧第一图像的位姿信息，可以更简便更高效地得到单帧第一图像的位姿信息。
可见，通过步骤SA1～SA3，针对背对相机拍摄的单帧第一图像，也可以准确地得到单帧第一图像中目标对象的位姿信息。
应当理解，步骤SA1～SA3的方法也可以适用于目标对象的正面图像，例如，针对人脸正面图像，可以得到人脸正面图像的位姿信息，其中，第一关键点，可以通过交互的方式手动标注，也可以对人脸正面图像进行特征点检测(例如包括基于深度学习方法等)，得到第一关键点。
图3示出根据本公开实施例的投影效果的示意图，图3以人头为目标对象，分别示出了两个不同拍摄角度下单帧第一图像的投影效果。如图3所示，三角形的10个点代表在单帧第一图像中标注的10个第一关键点，圆形的10个点代表三维的预设关键点(如图2左图10个点)投影到单帧第一图像的第一投影点。根据图3可知，单帧第一图像中标注的第一关键点与第一投影点可以重合或相逼近，说明本公开实施例的位姿检测方法，可以较准确的求解关于后脑勺图像的位姿信息。
上文介绍了针对单相机拍摄的单帧背面人头图像的场景的位姿检测，下面针对实验场景下相机阵列拍摄的多帧背面人头图像的场景，对本公开实施例的位姿检测方法进行说明。
其中，实验场景为人工布置的场景，可包括经过人工布置的实验室，使该场景下拍摄的图像的背景是人为设定的，例如背景是纯白色。
在一种可能的实现方式中，在步骤S1中，在图像采集设备为相机阵列的情况下，可获取待处理的至少一帧第一图像，所述至少一帧第一图像可包括：多帧第一图像以及多帧第二图像，所述第二图像为目标对象正面的二维图像。
在步骤S1确定了多帧第一图像以及多帧第二图像，可在步骤S2中根据多帧第一图像，以及由多帧第二图像三维重建的预设点云，确定每帧第一图像的目标对象的位姿信息，步骤S2可包括：
在步骤SB2中，对所述多帧第二图像进行三维重建，得到所述预设点云，以及对所述多帧第一图像进行三维重建，得到第一点云；
在步骤SB3中，根据所述预设点云和所述第一点云的相对关系，确定多帧第一图像中每帧第一图像中目标对象的位姿信息。
举例来说，相机阵列(Camera Arrays)可利用不同空间位置的多个相机来采集不同视角的图像。当每个相机之间的距离都比较大时，整个相机阵列可以看成是一个多中心投影相机(Multiple-Center-of-Projection Camera)，可以得到目标对象的多视角信息。
例如，假设相机阵列包括了M个相机，对目标对象的一次拍摄可以采集M帧不同视角的图像，本公开对相机阵列包括的相机数量M不做限制。
应当理解，在步骤S1中，可以使用相同的相机阵列拍摄多帧第一图像以及多帧第二图像，也可以使用不同的相机阵列分别拍摄多帧第一图像和多帧第二图像，获取的第一图像的帧数与获取的第二图像的帧数可以相同，也可以不同，本公开对第一图像帧数、 以及第二图像帧数的数量不作限制。为了便于下文的说明，可以使用一个包括了M个(例如30)相机阵列获取M帧第一图像和M帧第二图像。
在一种可能的实现方式中，在图像采集设备为相机阵列的情况下，通过相机阵列获取M帧第一图像以及M帧第二图像之前(即在步骤S1之前)，可以在步骤SB1中，根据所述相机阵列采集的棋盘格图像，标定相机阵列的内参数和外参数，所述内参数包括焦距或像素的至少一个，所述外参数包括相机位置或相机旋转角度的至少一个。
其中，所述棋盘格图像表示具有棋盘格图案的图像，棋盘格图案也即由两种颜色(例如黑白)相间的方格构成的图像。例如，棋盘格图像可以由2行×3列方格构成，第1行第1列、第1行第3列和第2行第2列为黑色方格，第1行第2列、第2行第1列和第2行第3列为白色方格，本公开对棋盘格的行数、列数、颜色、以及每个方格的边长不做限制。
对相机阵列进行标定，可确定三维空间中目标对象表面某点的三维几何位置与其在图像中对应点之间的相互关系，也即世界坐标到像素坐标的映射。对相机阵列的标定，可以对相机阵列包括的M个相机同时求解标定参数，获取相机阵列的标定参数trans_camera。
标定参数trans_camera可包括内参数和外参数，外参数用于确定现实世界三维空间中目标对象表面某点的几何位置与其在摄像机坐标中对应点的相互关系，也即世界坐标到摄像机坐标的映射，可包括相机位置，相机旋转角度等；内参数用于确定摄像机坐标下目标对象表面某点是如何继续经过摄像机的镜头、并通过针孔成像和电子转化而成为图像中的像素点，也即摄像机坐标到图像坐标的映射，可包括与相机自身特性相关的参数，比如焦距、像素等。
对相机阵列进行标定所采用的标定方法可包括：传统相机标定法、主动视觉相机标定方法、相机自标定法、零失真相机标定法等，本公开对标定方法不做限制。
为了更便捷地对相机阵列进行高精度标定，本公开的实施例例如可以使用单平面棋盘格相机标定方法，该方法可以减少或克服传统相机标定法需要的高精度标定物的缺点，仅使用一个打印出来的棋盘格就可以实现相机的标定，同时单平面棋盘格相机标定方法相对于相机自标定法而言，可以提高精度，并且更便于操作。
单平面棋盘格相机标定方法可基于单平面棋盘格进行相机标定，用于标定的棋盘格是三维场景(世界坐标)中的一个平面，对其拍摄的二维图像(图像坐标)是另一个平面，根据两个平面的对应点的坐标，可以求解相机的标定参数trans_camera。
例如，假设相机阵列由相机1～相机M构成，相机1～相机M对应的标定参数为trans_camera[1]～trans_camera[M]，针对相机阵列中的任一相机i(i＝1～M)，其标定参数trans_camera[i]可包括由矩阵A构成的内参数，以及由矩阵[R T]构成的外参数，假设三维的世界坐标系下的齐次坐标
Figure PCTCN2022134852-appb-000002
二维的图像坐标系下的齐次坐标
Figure PCTCN2022134852-appb-000003
标定用的棋盘格平面到图像平面的单应性可以表示为：
Figure PCTCN2022134852-appb-000004
其中，
Figure PCTCN2022134852-appb-000005
在公式(3)中，s代表世界坐标系到图像坐标系的尺度因子，在内参数A表示的矩阵中，为α和β代表图像坐标系下每个像素在x轴方向和y轴方向上的物理尺寸，c代表图像物理坐标的畸变参数，u 0和v 0代表图像原点；在外参数[R T]中，R代表世界坐标系转换到相机坐标系的旋转变换，T代表世界坐标系转换到相机坐标系的平移变换。
根据公式(3)，通过从不同角度拍摄的多张标定板图像，可以求解出相机的内参数和外参数。图4示出根据本公开实施例的标定板的示意图，如图4所示，可采用单平面棋盘格相机标定方法，使用棋盘格计算得到相机的标定参数trans_camera，其中，标定板可以是12×9的棋盘格，每个棋盘格的大小可以是边长10mm的正方形。应当理解，图4所示的棋盘格标定板仅做示例，本公开对标定板具有棋盘格的行数、列数、以及每 个棋盘格的大小不做限制。
因此，通过步骤SB1对相机阵列的标定，可以得到相机阵列的标定参数(内参数+外参数)，可用于确定相机阵列中每个相机的图像坐标系与世界坐标系的关系。
在步骤SB1对相机阵列标定之后，可以在步骤S1使用该相机阵列对目标对象进行图像采集。假设该相机阵列包括了M个设置在不同角度的相机，可以在目标对象的正面(例如人脸)背对相机阵列的情况下，获取M帧第一图像，以及在目标对象的正面(例如人脸)朝向相机阵列的情况下，获取M帧第二图像。应当理解，此处仅以M帧第一图像和M帧第二图像作为示例，第一图像的帧数与第二图像的帧数也可以不同，本公开对此不作限制。
在步骤SB2中，可以对M帧第二图像进行三维重建，得到预设点云，以及对所述M帧第一图像进行三维重建，得到第一点云。其中，预设点云为目标对象正面的三维点云，可以作为位姿检测的参考基准，第一点云为目标对象背面的三维点云。
应当理解，针对相机阵列采集的多帧图像对目标对象进行三维重建的方法，可以包括：基于传统方法比如双目视觉的重建方法、基于模型比如三维形变模型(3D Morphable Model)的重建方法、基于卷积神经网络的重建方法、基于三维建模软件的重建方法等，本公开的实施例对具体的三维重建方法不做限制。
在一种可能的实现方式中，步骤SB2可包括：根据所述内参数和所述外参数，对所述多帧第二图像进行三维重建，得到所述预设点云，以及根据所述内参数和所述外参数，对所述多帧第一图像进行三维重建，得到所述第一点云。
例如，可以利用预设的三维重建网络，将所述内参数和所述外参数、所述M帧第二图像输入该三维重建网络，生成预设点云，以及，将所述内参数和所述外参数、所述M帧第一图像输入该三维重建网络，生成预设点云。
其中，预设的三维重建网络也即训练好的神经网络，该三维重建网络经过大量的样本训练(学习)后，能够基于输入的内参数、外参数、多帧图像生成三维点云，本公开对三维重建网络的网络结构和网络训练方法不作限制。
或者，将所述内参数和所述外参数、所述M帧第二图像输入三维建模软件，得到所述预设点云，以及将所述内参数和所述外参数、所述M帧第一图像输入三维建模软件，得到所述第一点云。其中，三维建模软件可以通过相关技术实现，本公开对三维建模软件不做限制。
图5示出根据本公开实施例的预设点云获取方式的示意图，如图5所示，可以将步骤SB1得到的相机阵列的标定参数(内参数+外参数)，以及步骤SB1得到的M帧(例如M＝30)第二图像输入三维建模软件，得到如图5右侧所示的人头正面三维密集点云，即预设点云whole_scan。
相应的，图6示出根据本公开实施例的第一点云获取方式的示意图。如图6所示，可以将步骤SB1得到的相机阵列的标定参数(内参数+外参数)，以及步骤SB1得到的M帧(例如M＝30)第一图像输入三维建模软件，得到如图6右侧所示的背面三维密集点云，即第一点云part_scan。
其中，由于图6左侧的30帧第一图像是背面和侧面的人头图像，再加上头发的干扰，重建的第一点云part_scan容易出现残缺。
在步骤SB2得到预设点云whole_scan和第一点云part_scan，可在步骤SB3中，根据预设点云whole_scan和第一点云part_scan的相对关系，确定每帧第一图像中目标对象的位姿信息。
示例性的，所述预设点云和所述第一点云的相对关系包括所述预设点云与所述第一点云之间的位姿变换信息。
举例来说，可对所述预设点云whole_scan与所述第一点云part_scan进行对齐处理，确定所述预设点云与所述第一点云之间的相对关系。
举例来说，可根据标注的预设关键点确定预设点云whole_scan的初始位姿，可根据第二关键点确定第一点云part_scan的初始位姿。
可基于预设点云whole_scan的初始位姿和第一点云part_scan的初始位姿对预设点云whole_scan与所述第一点云part_scan进行对齐处理，例如，可以在保持预设点云whole_scan的初始位姿不变的情况下，改变(例如旋转、缩放、平移中的至少一种)第一点云part_scan的位姿，使预设点云whole_scan与第一点云part_scan对齐；或者，也可以在保持第一点云part_scan的初始位姿不变的情况下，改变预设点云whole_scan的位姿，使预设点云whole_scan的位姿与第一点云part_scan的初始位姿对齐；或者，还可以同时改变预设点云whole_scan的初始位姿和第一点云part_scan的初始位姿，以使两者对齐。本公开对此不作限制。
可对预设点云whole_scan与所述第一点云part_scan进行对齐处理，可使用迭达最近点(iteration closest points，ICP)方法，确定预设点云whole_scan与第一点云part_scan的位姿变换信息trans_scan。
例如，假设预设点云whole_scan的位姿变换信息trans_scan(例如包括缩放变换信息scale_2、旋转变换信息rotation_2、以及平移变换信息translation_2中的至少一种)，可表示为：trans_scan＝scale_2*rotation_2*whole_scan+translation_2。
在预设点云whole_scan的位姿变换信息trans_scan与第一点云part_scan两者的差值最小(预设点云whole_scan的位姿变换信息trans_scan无限逼近第一点云part_scan，也即预设点云whole_scan与第一点云part_scan对齐，如图7中间位置处的图所示)情况下，记录预设点云whole_scan与第一点云part_scan的位姿变换信息trans_scan。
通过这样方式，有利于准确地获取所述预设点云与所述第一点云之间的相对关系。
在确定了预设点云whole_scan与第一点云part_scan之间的相对关系，可在步骤SB3中，根据预设点云whole_scan和第一点云part_scan的相对关系，确定每帧第一图像中目标对象的位姿信息。
在一种可能的实现方式中，步骤SB3，可包括：
在步骤SB31’中，根据相机阵列的内参数和外参数、确定所述预设点云与所述第一点云之间的位姿变换信息，确定所述预设点云与每帧第一图像之间的位姿变换信息；
在步骤SB32’中，根据所述预设点云与每帧第一图像之间的位姿变换信息，确定所述多帧第一图像中每帧第一图像中目标对象的位姿信息。
举例来说，在步骤SB31’中，根据步骤SB1获取的相机阵列的标定参数trans_camera，包括相机阵列的内参数和外参数、以及预设点云whole_scan与第一点云part_scan的位姿变换信息trans_scan，可确定预设点云whole_scan相对于M帧第一图像的位姿变换信息。
在步骤SB32’中，可分别根据预设点云whole_scan相对于M帧第一图像的位姿变换信息，将预设点云whole_scan分别投影至各第一图像，确定每帧第一图像对应目标对象的位姿信息，该位姿信息可包括缩放信息，旋转信息和平移信息中的至少一种。
通过这种方式，可确定多帧第一图像中每帧第一图像中目标对象的位姿信息，该方法简单便捷。
然而在上述过程中，预设点云whole_scan是杂乱点序(没有规则拓扑)，且拥有噪声，直接根据预设点云whole_scan确定多帧第一图像中每帧第一图像中目标对象的位姿信息，容易降低位姿检测的准确性。
为了进一步提高位姿检测的准确性，可以将预设点云whole_scan统一为具有规则拓扑的网格数据(mesh数据)，以提高位姿检测的准确性。下面介绍基于网格数据(mesh数据)，确定多帧第一图像中每帧第一图像中目标对象的位姿信息。
在一种可能的实现方式中，对所述预设点云进行点云注册处理，确定所述预设点云对应的网格数据，以及所述网格数据与所述预设点云之间的位姿变换信息。
可基于点云注册方法，例如非刚性最近点迭达法(Nonrigid Iteration Closest Points)，点云注册方法用于将没有规则拓扑的预设点云whole_scan(如图5右图)统一为具有规则拓扑的网格数据mesh(如图2左图所示的光滑平面表示)。例如，可对初始状态的网格数据mesh和预设点云whole_scan进行非刚性配准，找到初始状态的网格数据mesh和预设点云whole_scan之间的映射，将初始状态的网格数据mesh进行变形以拟合预设点云whole_scan，同时保持网格数据mesh自身的拓扑结构。
在得到了网格数据mesh，可以求解预设点云whole_scan与网格数据mesh的位姿变换信息trans_mesh。
例如，假设网格数据mesh的位姿变换信息trans_mesh(例如包括缩放变换信息scale_1、旋转变换信息rotation_1、以及平移变换信息translation_1)，可表示为：trans_mesh＝scale_1*rotation_1*mesh+translation_1。
在网络数据mesh的位姿变换信息trans_mesh与预设点云whole_scan两者的差值最小(网络数据mesh的位姿变换信息trans_mesh无限逼近预设点云whole_scan，也即预设点云whole_scan与网格数据mesh对齐，如图7左图所示)的情况下，记录网格数据mesh与预设点云whole_scan的位姿变换信息trans_mesh。
在确定了网格数据与预设点云的位姿变换信息的过程中，还可以同步获取所述预设点云和所述第一点云的相对关系，所述相对关系包括所述预设点云与所述第一点云之间的位姿变换信息。具体预设点云和第一点云相对关系的确定方式可参考上文，此处不再赘叙。
在获取到网格数据与预设点云的位姿变换信息，以及所述预设点云与所述第一点云之间的相对关系，可在步骤SB3确定每帧第一图像中目标对象的位姿信息。
在一种可能的实现方式中，步骤SB3，可包括：
在步骤SB31中，根据所述网格数据与所述预设点云之间的位姿变换信息，所述预设点云与所述第一点云之间的位姿变换信息，得到所述网格数据与所述第一点云之间的位姿变换信息；
在步骤SB32中，根据相机阵列的内参数和外参数、所述网格数据与所述第一点云之间的位姿变换信息，确定所述网格数据与每帧第一图像之间的位姿变换信息；
在步骤SB33中，根据所述网格数据与每帧第一图像之间的位姿变换信息，确定多帧第一图像中每帧第一图像中目标对象的位姿信息。
举例来说，图7示出根据本公开实施例的网格数据和第一点云对齐的示意图。步骤SB31～SB33介绍了网格数据和第一点云对齐的方法，具体如下：
在步骤SB31中，根据步骤网格数据mesh与预设点云whole_scan的位姿变换信息trans_mesh，以及预设点云whole_scan与第一点云part_scan的位姿变换信息trans_scan，得到网格数据mesh到第一点云part_scan的位姿变换信息trans_part；
例如，网格数据mesh与第一点云part_scan的位姿变换信息trans_part为：scale_2*rotation_2*(scale_1*rotation_1*mesh+translation_1)+translation_2，可以使网格数据mesh与第一点云part_scan对齐，如图7右图所示。
在步骤SB32中，根据步骤SB1获取的相机阵列的标定参数trans_camera，包括相机阵列的内参数和外参数、以及网格数据mesh与第一点云part_scan的位姿变换信息trans_part，可确定网格数据mesh相对于M帧第一图像的位姿变换信息trans_end；
在步骤SB33中，可分别根据所述网格数据mesh相对于M帧第一图像的位姿变换信息trans_end，将网格数据mesh分别投影至各第一图像，确定每帧第一图像对应目标对象的位姿信息，该位姿信息可包括缩放信息，旋转信息和平移信息。图8示出根据本公开实施例的投影效果的示意图。如图8所示，网格数据mesh可渲染到每帧人头背面图像，根据本公开实施例可以准确地得到人头背面图像的位姿信息。
应当理解，在上述过程中，由于在预设点云whole_scan和第一点云part_scan均包 括了大量的点，直接分析这两个点云，需要进行大规模的计算，会消耗大量的计算资源。为了提高计算效率，可以在预设点云whole_scan和第一点云part_scan上分别标注关键点，可根据关键点的对齐方式，实现点云的对齐，进而基于关键点确定每帧第一图像中目标对象的位姿信息。
在一种可能的实现方式中，所述对所述预设点云与所述第一点云进行对齐处理，确定所述预设点云与所述第一点云之间的相对关系，可包括：在所述预设点云的人脸区域中，生成包含N个预设关键点的标注信息；在所述第一点云的人脸区域中，生成包含K个第二关键点的标注信息，K为不大于N的整数；对所述预设点云的N个关键点与所述第一点云的K个第二关键点进行对齐处理，确定预设点云与所述第一点云之间的相对关系。
举例来说，如图5右侧的人头正面的三维密集点云(即预设点云whole_scan)所示，可以在人脸区域标注N个预设关键点，图5示出了预设关键点为10个的情况，可分别对应左/右外侧眼角点，左/右内侧眼角点，鼻尖点，左/右嘴角点，下巴中心点，左/右侧耳根点。
如图6右侧人头的三维密集点云(即第一点云part_scan)所示，可以在人脸区域标注K个第二关键点，图6示出了预设关键点的情况，由于图6重建的第一点云part_scan，部分人脸区域是缺失的，可以忽略缺失区域的关键点(例如可以忽略右外侧眼角点，右眼内侧眼角点，右嘴角点，右侧耳根点这4个关键点)，在非缺失区域标注6个(K＝6)第二关键点，可分别对应左眼外侧眼角点，左眼内侧眼角点，鼻尖点，左嘴角点，下巴中心点，左侧耳根点。
其中，在人脸区域获取N个预设关键点和/或K个第二关键点，可以通过接收的用户标注操作进行标注，也可以对人脸进行特征点检测(例如包括基于深度学习方法等)，得到标注点，本公开对生成标注信息的方法不作限制。
应当理解，图5和图6仅做示意，可在与预设关键点对应部位处生成第二关键点的标注信息，其对应部位处的关键点拥有相同的点序，本公开对预设关键点和第二关键点的数量，以及预设关键点和第二关键点的具体标注位置不做限制。
在生成了关键点的标注信息之后，可以先从预设点云的N个预设关键点中，选择与K个第二关键对应的预设关键点，再根据选中的K个预设关键点和K个第二关键点进行对齐处理，确定预设点云与所述第一点云之间的相对关系。具体的对齐处理可参考上文对齐方式，此处不再赘叙。
通过这种方式，可以提高计算效率，节约计算资源。
通过步骤SB1～SB3，针对实验室环境下多相机阵列拍摄的多帧背面人头图像的场景，能够准确的确定各帧第一图像中目标对象的位姿信息。
因此，根据本公开的实施例，针对一些无法准确提供明确语义部位的目标对象背面图像，也可以准确地确定出图像中目标对象相对于目标对象的三维模型(预设点云)，不仅能准确实现背对相机拍摄的单帧图像的位姿(步骤SA1～SA3)，还能准确实现背对相机阵列设备拍摄的多帧图像的位姿(步骤SB1～SB3)，有效解决目标对象全方位的位姿检测，不管目标对象的面部朝向与拍摄方向的角度呈现什么角度，均能实现准确地位姿检测。
可以理解，本公开提及的上述各个方法实施例，在不违背原理逻辑的情况下，均可以彼此相互结合形成结合后的实施例，限于篇幅，本公开不再赘述。本领域技术人员可以理解，在具体实施方式的上述方法中，各步骤的具体执行顺序应当以其功能和可能的内在逻辑确定。
此外，本公开还提供了位姿检测装置、电子设备、计算机可读存储介质、程序，上述均可用来实现本公开提供的任一种位姿检测方法，相应技术方案和描述和参见方法部分的相应记载，不再赘述。
图9示出根据本公开实施例的位姿检测装置的框图，如图9所示，所述装置包括：
获取模块91，用于获取至少一帧第一图像，其中，所述至少一帧第一图像包括图像采集设备采集的目标对象背面的至少一帧二维图像；
确定模块92，用于根据所述至少一帧第一图像，以及预设点云，确定所述至少一帧第一图像的每个中所述目标对象的位姿信息，其中，所述预设点云为所述目标对象正面的三维点云。
在一种可能的实现方式中，所述预设点云包括N个预设关键点，N为大于1的整数，在所述图像采集设备为单相机的情况下，所述至少一帧第一图像包括所述单相机拍摄的单帧第一图像；所述确定模块92用于：在所述单帧第一图像的目标对象区域，生成包含N个第一关键点的标注信息；将所述预设点云投影至所述单帧第一图像，确定N个第一关键点对应的N个第一投影点；对所述N个第一投影点与所述N个预设关键点进行拟合处理，得到所述单帧第一图像中目标对象的位姿信息。
在一种可能的实现方式中，在图像采集设备为相机阵列的情况下，所述至少一帧第一图像包括：多帧第一图像以及多帧第二图像，所述多帧第二图像为所述目标对象正面的多帧二维图像；所述确定模块92用于：对所述多帧第二图像进行三维重建，得到所述预设点云，以及对所述多帧第一图像进行三维重建，得到第一点云；根据所述预设点云和所述第一点云的相对关系，确定所述多帧第一图像中每帧第一图像中目标对象的位姿信息。
在一种可能的实现方式中，所述获取模块91，用于在图像采集设备为相机阵列的情况下，在所述获取至少一帧第一图像之前，根据所述相机阵列采集的棋盘格图像，标定所述相机阵列的内参数和外参数，所述内参数包括焦距或像素中的至少一个，所述外参数包括相机位置或相机旋转角度中的至少一个；所述确定模块92用于：根据所述内参数和所述外参数，对所述多帧第二图像进行三维重建，得到所述预设点云；以及，根据所述内参数和所述外参数，对所述多帧第一图像进行三维重建，得到所述第一点云。
在一种可能的实现方式中，所述预设点云和所述第一点云的相对关系包括所述预设点云与所述第一点云之间的位姿变换信息，所述根据所述预设点云和所述第一点云的相对关系，确定所述多帧第一图像中每帧第一图像中目标对象的位姿信息，包括：根据相机阵列的内参数和外参数、以及所述预设点云与所述第一点云之间的位姿变换信息，确定所述预设点云与所述多帧第一图像中每帧第一图像之间的位姿变换信息；根据所述预设点云与所述多帧第一图像中每帧第一图像之间的位姿变换信息，确定所述多帧第一图像中每帧第一图像中所述目标对象的位姿信息。
在一种可能的实现方式中，所述预设点云和所述第一点云的相对关系包括所述预设点云与所述第一点云之间的位姿变换信息，所述确定模块92还用于：对所述预设点云进行点云注册处理，确定所述预设点云对应的网格数据，以及所述网格数据与所述预设点云之间的位姿变换信息；所述根据所述预设点云和所述第一点云的相对关系，确定所述多帧第一图像中每帧第一图像中目标对象的位姿信息，包括：根据所述网格数据与所述预设点云之间的位姿变换信息，以及所述预设点云与所述第一点云之间的位姿变换信息，确定所述网格数据与所述第一点云之间的位姿变换信息；根据相机阵列的内参数和外参数、所述网格数据与所述第一点云之间的位姿变换信息，确定所述网格数据与所述多帧第一图像中每帧第一图像之间的位姿变换信息；根据所述网格数据与所述多帧第一图像中每帧第一图像之间的位姿变换信息，确定所述多帧第一图像中每帧第一图像中所述目标对象的位姿信息。
在一种可能的实现方式中，所述确定模块92还用于：在所述对所述多帧第二图像进行三维重建，得到所述预设点云，以及对所述多帧第一图像进行三维重建，得到第一点云之后，对所述预设点云与所述第一点云进行对齐处理，确定所述预设点云与所述第一点云之间的相对关系。
在一种可能的实现方式中，所述对所述预设点云与所述第一点云进行对齐处理，确定所述预设点云与所述第一点云之间的相对关系，包括：在所述预设点云的人脸区域中，生成包含N个预设关键点的标注信息；在所述第一点云的人脸区域中，生成包含K个第二关键点的标注信息，K为不大于N的整数；对所述预设点云的N个关键点与所述第一点云的K个第二关键点进行对齐处理，确定预设点云与所述第一点云之间的相对关系。
在一种可能的实现方式中，所述目标对象包括人体的身体部位；所述人体的身体部位包括人脸的情况下，所述预设关键点包括人脸器官的关键点；所述人体的身体部位包括肢体的情况下，所述预设关键点包括肢体关键点。
在一些实施例中，本公开实施例提供的装置具有的功能或包含的模块可以用于执行上文方法实施例描述的方法，其具体实现可以参照上文方法实施例的描述，为了简洁，这里不再赘述。
本公开实施例还提出一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现上述方法。计算机可读存储介质可以是易失性或非易失性计算机可读存储介质。
本公开实施例还提出一种电子设备，包括：处理器；用于存储处理器可执行指令的存储器；其中，所述处理器被配置为调用所述存储器存储的指令，以执行上述方法。
本公开实施例还提供了一种计算机程序产品，包括计算机可读代码，或者承载有计算机可读代码的非易失性计算机可读存储介质，当所述计算机可读代码在电子设备的处理器中运行时，所述电子设备中的处理器执行上述方法。
电子设备可以被提供为终端、服务器或其它形态的设备。
图10示出根据本公开实施例的一种电子设备800的框图。例如，电子设备800可以是移动电话，计算机，数字广播终端，消息收发设备，游戏控制台，平板设备，医疗设备，健身设备，个人数字助理等终端。
参照图10，电子设备800可以包括以下一个或多个组件：处理组件802，存储器804，电源组件806，多媒体组件808，音频组件810，输入/输出(I/O)的接口812，传感器组件814，以及通信组件816。
处理组件802通常控制电子设备800的整体操作，诸如与显示，电话呼叫，数据通信，相机操作和记录操作相关联的操作。处理组件802可以包括一个或多个处理器820来执行指令，以完成上述的方法的全部或部分步骤。此外，处理组件802可以包括一个或多个模块，便于处理组件802和其他组件之间的交互。例如，处理组件802可以包括多媒体模块，以方便多媒体组件808和处理组件802之间的交互。
存储器804被配置为存储各种类型的数据以支持在电子设备800的操作。这些数据的示例包括用于在电子设备800上操作的任何应用程序或方法的指令，联系人数据，电话簿数据，消息，图片，视频等。存储器804可以由任何类型的易失性或非易失性存储设备或者它们的组合实现，如静态随机存取存储器(SRAM)，电可擦除可编程只读存储器(EEPROM)，可擦除可编程只读存储器(EPROM)，可编程只读存储器(PROM)，只读存储器(ROM)，磁存储器，快闪存储器，磁盘或光盘。
电源组件806为电子设备800的各种组件提供电力。电源组件806可以包括电源管理系统，一个或多个电源，及其他与为电子设备800生成、管理和分配电力相关联的组件。
多媒体组件808包括在所述电子设备800和用户之间的提供一个输出接口的屏幕。在一些实施例中，屏幕可以包括液晶显示器(LCD)和触摸面板(TP)。如果屏幕包括触摸面板，屏幕可以被实现为触摸屏，以接收来自用户的输入信号。触摸面板包括一个或多个触摸传感器以感测触摸、滑动和触摸面板上的手势。所述触摸传感器可以不仅感测触摸或滑动动作的边界，而且还检测与所述触摸或滑动操作相关的持续时间和压力。 在一些实施例中，多媒体组件808包括一个前置摄像头和/或后置摄像头。当电子设备800处于操作模式，如拍摄模式或视频模式时，前置摄像头和/或后置摄像头可以接收外部的多媒体数据。每个前置摄像头和后置摄像头可以是一个固定的光学透镜系统或具有焦距和光学变焦能力。
音频组件810被配置为输出和/或输入音频信号。例如，音频组件810包括一个麦克风(MIC)，当电子设备800处于操作模式，如呼叫模式、记录模式和语音识别模式时，麦克风被配置为接收外部音频信号。所接收的音频信号可以被进一步存储在存储器804或经由通信组件816发送。在一些实施例中，音频组件810还包括一个扬声器，用于输出音频信号。
I/O接口812为处理组件802和外围接口模块之间提供接口，上述外围接口模块可以是键盘，点击轮，按钮等。这些按钮可包括但不限于：主页按钮、音量按钮、启动按钮和锁定按钮。
传感器组件814包括一个或多个传感器，用于为电子设备800提供各个方面的状态评估。例如，传感器组件814可以检测到电子设备800的打开/关闭状态，组件的相对定位，例如所述组件为电子设备800的显示器和小键盘，传感器组件814还可以检测电子设备800或电子设备800一个组件的位置改变，用户与电子设备800接触的存在或不存在，电子设备800方位或加速/减速和电子设备800的温度变化。传感器组件814可以包括接近传感器，被配置用来在没有任何的物理接触时检测附近物体的存在。传感器组件814还可以包括光传感器，如互补金属氧化物半导体(CMOS)或电荷耦合装置(CCD)图像传感器，用于在成像应用中使用。在一些实施例中，该传感器组件814还可以包括加速度传感器，陀螺仪传感器，磁传感器，压力传感器或温度传感器。
通信组件816被配置为便于电子设备800和其他设备之间有线或无线方式的通信。电子设备800可以接入基于通信标准的无线网络，如无线网络(Wi-Fi)、第二代移动通信技术(2G)、第三代移动通信技术(3G)、第四代移动通信技术(4G)、通用移动通信技术的长期演进(LTE)、第五代移动通信技术(5G)或它们的组合。在一个示例性实施例中，通信组件816经由广播信道接收来自外部广播管理系统的广播信号或广播相关信息。在一个示例性实施例中，所述通信组件816还包括近场通信(NFC)模块，以促进短程通信。例如，在NFC模块可基于射频识别(RFID)技术，红外数据协会(IrDA)技术，超宽带(UWB)技术，蓝牙(BT)技术和其他技术来实现。
在示例性实施例中，电子设备800可以被一个或多个应用专用集成电路(ASIC)、数字信号处理器(DSP)、数字信号处理设备(DSPD)、可编程逻辑器件(PLD)、现场可编程门阵列(FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行上述方法。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器804，上述计算机程序指令可由电子设备800的处理器820执行以完成上述方法。
本公开涉及增强现实领域，通过获取现实环境中的目标对象的图像信息，进而借助各类视觉相关算法实现对目标对象的相关特征、状态及属性进行检测或识别处理，从而得到与具体应用匹配的虚拟与现实相结合的AR效果。示例性的，目标对象可涉及与人体相关的脸部、肢体、手势、动作等，或者与物体相关的标识物、标志物，或者与场馆或场所相关的沙盘、展示区域或展示物品等。视觉相关算法可涉及视觉定位、SLAM、三维重建、图像注册、背景分割、对象的关键点提取及跟踪、对象的位姿或深度检测等。具体应用不仅可以涉及跟真实场景或物品相关的导览、导航、讲解、重建、虚拟效果叠加展示等交互场景，还可以涉及与人相关的特效处理，比如妆容美化、肢体美化、特效展示、虚拟模型展示等交互场景。可通过卷积神经网络，实现对目标对象的相关特征、状态及属性进行检测或识别处理。上述卷积神经网络是基于深度学习框架进行模型训练 而得到的网络模型。
图11示出根据本公开实施例的一种电子设备1900的框图。例如，电子设备1900可以被提供为一服务器。参照图11，电子设备1900包括处理组件1922，其进一步包括一个或多个处理器，以及由存储器1932所代表的存储器资源，用于存储可由处理组件1922的执行的指令，例如应用程序。存储器1932中存储的应用程序可以包括一个或一个以上的每一个对应于一组指令的模块。此外，处理组件1922被配置为执行指令，以执行上述方法。
电子设备1900还可以包括一个电源组件1926被配置为执行电子设备1900的电源管理，一个有线或无线网络接口1950被配置为将电子设备1900连接到网络，和一个输入输出(I/O)接口1958。电子设备1900可以操作基于存储在存储器1932的操作系统，例如微软服务器操作系统(Windows ServerTM)，苹果公司推出的基于图形用户界面操作系统(Mac OS XTM)，多用户多进程的计算机操作系统(UnixTM),自由和开放原代码的类Unix操作系统(LinuxTM)，开放原代码的类Unix操作系统(FreeBSDTM)或类似。
在示例性实施例中，还提供了一种非易失性计算机可读存储介质，例如包括计算机程序指令的存储器1932，上述计算机程序指令可由电子设备1900的处理组件1922执行以完成上述方法。
本公开可以是系统、方法和/或计算机程序产品。计算机程序产品可以包括计算机可读存储介质，其上载有用于使处理器实现本公开的各个方面的计算机可读程序指令。
计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是(但不限于)电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身，诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如，通过光纤电缆的光脉冲)、或者通过电线传输的电信号。
这里所描述的计算机可读程序指令可以从计算机可读存储介质下载到各个计算/处理设备，或者通过网络、例如因特网、局域网、广域网和/或无线网下载到外部计算机或外部存储设备。网络可以包括铜传输电缆、光纤传输、无线传输、路由器、防火墙、交换机、网关计算机和/或边缘服务器。每个计算/处理设备中的网络适配卡或者网络接口从网络接收计算机可读程序指令，并转发该计算机可读程序指令，以供存储在各个计算/处理设备中的计算机可读存储介质中。
用于执行本公开操作的计算机程序指令可以是汇编指令、指令集架构(ISA)指令、机器指令、机器相关指令、微代码、固件指令、状态设置数据、或者以一种或多种编程语言的任意组合编写的源代码或目标代码，所述编程语言包括面向对象的编程语言—诸如Smalltalk、C++等，以及常规的过程式编程语言—诸如“C”语言或类似的编程语言。计算机可读程序指令可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络—包括局域网(LAN)或广域网(WAN)—连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。在一些实施例中，通过利用计算机可读程序指令的状态信息来个性化定制电子电路，例如可编程逻辑电路、现场可编程门阵列(FPGA)或可编程逻辑阵列(PLA)，该电子电路可以执行计算机可 读程序指令，从而实现本公开的各个方面。
这里参照根据本公开实施例的方法、装置(系统)和计算机程序产品的流程图和/或框图描述了本公开的各个方面。应当理解，流程图和/或框图的每个方框以及流程图和/或框图中各方框的组合，都可以由计算机可读程序指令实现。
这些计算机可读程序指令可以提供给通用计算机、专用计算机或其它可编程数据处理装置的处理器，从而生产出一种机器，使得这些指令在通过计算机或其它可编程数据处理装置的处理器执行时，产生了实现流程图和/或框图中的一个或多个方框中规定的功能/动作的装置。也可以把这些计算机可读程序指令存储在计算机可读存储介质中，这些指令使得计算机、可编程数据处理装置和/或其他设备以特定方式工作，从而，存储有指令的计算机可读介质则包括一个制造品，其包括实现流程图和/或框图中的一个或多个方框中规定的功能/动作的各个方面的指令。
也可以把计算机可读程序指令加载到计算机、其它可编程数据处理装置、或其它设备上，使得在计算机、其它可编程数据处理装置或其它设备上执行一系列操作步骤，以产生计算机实现的过程，从而使得在计算机、其它可编程数据处理装置、或其它设备上执行的指令实现流程图和/或框图中的一个或多个方框中规定的功能/动作。
附图中的流程图和框图显示了根据本公开的多个实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或指令的一部分，所述模块、程序段或指令的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
该计算机程序产品可以具体通过硬件、软件或其结合的方式实现。在一个可选实施例中，所述计算机程序产品具体体现为计算机存储介质，在另一个可选实施例中，计算机程序产品具体体现为软件产品，例如软件开发包(Software Development Kit，SDK)等等。
以上已经描述了本公开的各实施例，上述说明是示例性的，并非穷尽性的，并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下，对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择，旨在最好地解释各实施例的原理、实际应用或对市场中的技术的改进，或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。

Claims (12)

  1. 一种位姿检测方法，包括：
    获取至少一帧第一图像，其中，所述至少一帧第一图像包括图像采集设备采集的目标对象背面的至少一帧二维图像；
    根据所述至少一帧第一图像，以及预设点云，确定所述至少一帧第一图像的每个中所述目标对象的位姿信息，其中，所述预设点云为所述目标对象正面的三维点云。
  2. 根据权利要求1所述的方法，其特征在于，所述预设点云包括N个预设关键点，N为大于1的整数，在所述图像采集设备为单相机的情况下，所述至少一帧第一图像包括所述单相机拍摄的单帧第一图像；
    所述根据所述至少一帧第一图像，以及预设点云，确定所述至少一帧第一图像的每个中所述目标对象的位姿信息，包括：
    在所述单帧第一图像的目标对象区域，生成包含N个第一关键点的标注信息；
    将所述预设点云投影至所述单帧第一图像，确定所述N个第一关键点对应的N个第一投影点；
    对所述N个第一投影点与所述N个预设关键点进行拟合处理，得到所述单帧第一图像中所述目标对象的位姿信息。
  3. 根据权利要求1所述的方法，其特征在于，在图像采集设备为相机阵列的情况下，所述至少一帧第一图像包括：多帧第一图像以及多帧第二图像，所述多帧第二图像为所述目标对象正面的多帧二维图像；
    所述根据所述至少一帧第一图像，以及预设点云，确定所述至少一帧第一图像的每个中所述目标对象的位姿信息，包括：
    对所述多帧第二图像进行三维重建，得到所述预设点云，以及对所述多帧第一图像进行三维重建，得到第一点云；
    根据所述预设点云和所述第一点云的相对关系，确定所述多帧第一图像中每帧第一图像中所述目标对象的位姿信息。
  4. 根据权利要求3所述的方法，其特征在于，在所述图像采集设备为相机阵列的情况下，在所述获取至少一帧第一图像之前，包括：
    根据所述相机阵列采集的棋盘格图像，标定所述相机阵列的内参数和外参数，所述内参数包括焦距或像素中的至少一个，所述外参数包括相机位置或相机旋转角度中的至少一个；
    所述对所述多帧第二图像进行三维重建，得到所述预设点云，包括：根据所述内参数和所述外参数，对所述多帧第二图像进行三维重建，得到所述预设点云；以及，
    所述对所述多帧第一图像进行三维重建，得到第一点云，包括：根据所述内参数和所述外参数，对所述多帧第一图像进行三维重建，得到所述第一点云。
  5. 根据权利要求3或4所述的方法，其特征在于，所述预设点云和所述第一点云的相对关系包括所述预设点云与所述第一点云之间的位姿变换信息，
    所述根据所述预设点云和所述第一点云的相对关系，确定所述多帧第一图像的每帧第一图像中所述目标对象的位姿信息，包括：
    根据所述相机阵列的内参数和外参数、以及所述预设点云与所述第一点云之间的位姿变换信息，确定所述预设点云与所述多帧第一图像中每帧第一图像之间的位姿变换信息；
    根据所述预设点云与所述多帧第一图像中每帧第一图像之间的位姿变换信息，确定所述多帧第一图像中每帧第一图像中所述目标对象的位姿信息。
  6. 根据权利要求3或4所述的方法，其特征在于，所述预设点云和所述第一点云的相对关系包括所述预设点云与所述第一点云之间的位姿变换信息，
    所述方法还包括：
    对所述预设点云进行点云注册处理，确定所述预设点云对应的网格数据，以及所述网格数据与所述预设点云之间的位姿变换信息；
    所述根据所述预设点云和所述第一点云的相对关系，确定所述多帧第一图像中每帧第一图像中所述目标对象的位姿信息，包括：根据所述网格数据与所述预设点云之间的位姿变换信息，以及所述预设点云与所述第一点云之间的位姿变换信息，确定所述网格数据与所述第一点云之间的位姿变换信息；
    根据所述相机阵列的内参数和外参数、所述网格数据与所述第一点云之间的所述位姿变换信息，确定所述网格数据与所述多帧第一图像中每帧第一图像之间的位姿变换信息；
    根据所述网格数据与所述多帧第一图像中每帧第一图像之间的位姿变换信息，确定所述多帧第一图像中每帧第一图像中所述目标对象的位姿信息。
  7. 根据权利要求3或4所述的方法，其特征在于，所述对所述多帧第二图像进行三维重建，得到所述预设点云，以及对所述多帧第一图像进行三维重建，得到第一点云之后，所述方法还包括：
    对所述预设点云与所述第一点云进行对齐处理，确定所述预设点云与所述第一点云之间的相对关系。
  8. 根据权利要求7所述的方法，其特征在于，所述对所述预设点云与所述第一点云进行对齐处理，确定所述预设点云与所述第一点云之间的相对关系，包括：
    在所述预设点云的人脸区域中，生成包含N个预设关键点的标注信息；
    在所述第一点云的人脸区域中，生成包含K个第二关键点的标注信息，K为不大于N的整数；
    对所述预设点云的N个关键点与所述第一点云的K个第二关键点进行对齐处理，确定所述预设点云与所述第一点云之间的相对关系。
  9. 根据权利要求2或8所述的方法，所述目标对象包括人体的身体部位；
    所述人体的身体部位包括人脸的情况下，所述预设关键点包括人脸器官的关键点；
    所述人体的身体部位包括肢体的情况下，所述预设关键点包括肢体关键点。
  10. 一种位姿检测装置，包括：
    获取模块，用于获取至少一帧第一图像，其中，所述至少一帧第一图像包括图像采集设备采集的目标对象背面的至少一帧二维图像；
    确定模块，用于根据所述至少一帧第一图像，以及预设点云，确定所述至少一帧第一图像的每个中所述目标对象的位姿信息，其中，所述预设点云为所述目标对象正面的三维点云。
  11. 一种电子设备，包括：
    处理器；
    用于存储处理器可执行指令的存储器；
    其中，所述处理器被配置为调用所述存储器存储的指令，以执行权利要求1至9中任意一项所述的方法。
  12. 一种计算机可读存储介质，其上存储有计算机程序指令，所述计算机程序指令被处理器执行时实现权利要求1至9中任意一项所述的方法。
PCT/CN2022/134852 2022-02-17 2022-11-29 位姿检测方法及装置、电子设备和存储介质 WO2023155532A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210147653.7A CN114529606A (zh) 2022-02-17 2022-02-17 位姿检测方法及装置、电子设备和存储介质
CN202210147653.7 2022-02-17

Publications (1)

Publication Number Publication Date
WO2023155532A1 true WO2023155532A1 (zh) 2023-08-24

Family

ID=81622069

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/134852 WO2023155532A1 (zh) 2022-02-17 2022-11-29 位姿检测方法及装置、电子设备和存储介质

Country Status (2)

Country Link
CN (1) CN114529606A (zh)
WO (1) WO2023155532A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117351306A (zh) * 2023-12-04 2024-01-05 齐鲁空天信息研究院 三维点云投影位姿求解器训练方法、确定方法及装置

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114529606A (zh) * 2022-02-17 2022-05-24 北京大甜绵白糖科技有限公司 位姿检测方法及装置、电子设备和存储介质
CN115661493B (zh) * 2022-12-28 2023-07-04 航天云机(北京)科技有限公司 一种对象位姿的确定方法及装置、设备及存储介质
CN116977417A (zh) * 2023-06-30 2023-10-31 北京百度网讯科技有限公司 位姿估计方法、装置、电子设备及存储介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2015215299A (ja) * 2014-05-13 2015-12-03 株式会社デンソー 対象物位置推定装置
CN112132107A (zh) * 2020-10-09 2020-12-25 腾讯科技（深圳）有限公司 图像处理方法、装置、电子设备及计算机可读存储介质
CN113689503A (zh) * 2021-10-25 2021-11-23 北京市商汤科技开发有限公司 目标对象的姿态检测方法、装置、设备及存储介质
CN113902853A (zh) * 2021-10-18 2022-01-07 深圳追一科技有限公司 一种人脸三维重建方法、装置、电子设备及存储介质
CN113902851A (zh) * 2021-10-18 2022-01-07 深圳追一科技有限公司 一种人脸三维重建方法、装置、电子设备及存储介质
CN113902852A (zh) * 2021-10-18 2022-01-07 深圳追一科技有限公司 一种人脸三维重建方法、装置、电子设备及存储介质
CN114529606A (zh) * 2022-02-17 2022-05-24 北京大甜绵白糖科技有限公司 位姿检测方法及装置、电子设备和存储介质

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2015215299A (ja) * 2014-05-13 2015-12-03 株式会社デンソー 対象物位置推定装置
CN112132107A (zh) * 2020-10-09 2020-12-25 腾讯科技（深圳）有限公司 图像处理方法、装置、电子设备及计算机可读存储介质
CN113902853A (zh) * 2021-10-18 2022-01-07 深圳追一科技有限公司 一种人脸三维重建方法、装置、电子设备及存储介质
CN113902851A (zh) * 2021-10-18 2022-01-07 深圳追一科技有限公司 一种人脸三维重建方法、装置、电子设备及存储介质
CN113902852A (zh) * 2021-10-18 2022-01-07 深圳追一科技有限公司 一种人脸三维重建方法、装置、电子设备及存储介质
CN113689503A (zh) * 2021-10-25 2021-11-23 北京市商汤科技开发有限公司 目标对象的姿态检测方法、装置、设备及存储介质
CN114529606A (zh) * 2022-02-17 2022-05-24 北京大甜绵白糖科技有限公司 位姿检测方法及装置、电子设备和存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117351306A (zh) * 2023-12-04 2024-01-05 齐鲁空天信息研究院 三维点云投影位姿求解器训练方法、确定方法及装置
CN117351306B (zh) * 2023-12-04 2024-03-22 齐鲁空天信息研究院 三维点云投影位姿求解器训练方法、确定方法及装置

Also Published As

Publication number Publication date
CN114529606A (zh) 2022-05-24

Similar Documents

Publication Publication Date Title
WO2023155532A1 (zh) 位姿检测方法及装置、电子设备和存储介质
US11756223B2 (en) Depth-aware photo editing
US10915998B2 (en) Image processing method and device
CN106664376B (zh) 增强现实设备和方法
JP5847924B2 (ja) 拡張現実表現のための二次元画像取込み
US20170372449A1 (en) Smart capturing of whiteboard contents for remote conferencing
CN109934931B (zh) 采集图像、建立目标物体识别模型的方法及装置
US11527014B2 (en) Methods and systems for calibrating surface data capture devices
TW201327472A (zh) 影像處理方法以及影像處理裝置
US9137461B2 (en) Real-time camera view through drawn region for image capture
JP2016537901A (ja) ライトフィールド処理方法
TWI752473B (zh) 圖像處理方法及裝置、電子設備和電腦可讀儲存媒體
WO2023051356A1 (zh) 一种虚拟对象的显示方法及装置、电子设备和存储介质
CN114445562A (zh) 三维重建方法及装置、电子设备和存储介质
WO2023273499A1 (zh) 深度检测方法及装置、电子设备和存储介质
WO2022142298A1 (zh) 关键点检测方法及装置、电子设备和存储介质
CN112991381B (zh) 图像处理方法及装置、电子设备和存储介质
CN106713740A (zh) 定位跟踪摄像方法与系统
WO2023273498A1 (zh) 深度检测方法及装置、电子设备和存储介质
CN103500471A (zh) 实现高分辨率增强现实系统的方法
CN114140536A (zh) 位姿数据处理方法、装置、电子设备及存储介质
CN114581525A (zh) 姿态确定方法及装置、电子设备和存储介质
WO2022016803A1 (zh) 视觉定位方法及装置、电子设备和计算机可读存储介质
CN112613447A (zh) 关键点检测方法及装置、电子设备和存储介质
WO2022151687A1 (zh) 合影图像生成方法、装置、设备、存储介质、计算机程序及产品

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22926836

Country of ref document: EP

Kind code of ref document: A1