WO2024060708A1 - 目标检测方法和装置 - Google Patents

目标检测方法和装置 Download PDF

Info

Publication number
WO2024060708A1
WO2024060708A1 PCT/CN2023/100388 CN2023100388W WO2024060708A1 WO 2024060708 A1 WO2024060708 A1 WO 2024060708A1 CN 2023100388 W CN2023100388 W CN 2023100388W WO 2024060708 A1 WO2024060708 A1 WO 2024060708A1
Authority
WO
WIPO (PCT)
Prior art keywords
network
target
position information
detected
deviation
Prior art date
Application number
PCT/CN2023/100388
Other languages
English (en)
French (fr)
Inventor
鲍慊
刘武
孙宇
梅涛
Original Assignee
北京京东尚科信息技术有限公司
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 北京京东尚科信息技术有限公司 filed Critical 北京京东尚科信息技术有限公司
Publication of WO2024060708A1 publication Critical patent/WO2024060708A1/zh

Links

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
    • 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
    • G06T17/05—Geographic models
    • 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

Definitions

  • Embodiments of the present disclosure relate to the field of computer technology, and specifically to target detection methods and devices.
  • Object detection is an important research direction in the field of computer vision, and it has good application prospects in fields such as intelligent transportation, intelligent assisted driving, and video surveillance. With the rapid development of deep learning, using deep learning to achieve target detection is one of the most studied target detection algorithms.
  • Embodiments of the present disclosure propose target detection methods and devices.
  • a target detection method includes: acquiring a two-dimensional image containing a target to be detected; inputting the image into a pre-trained target detection network to obtain spatial position information of the target to be detected , among which, the target detection network includes the main view position detection network, the bird's-eye view position detection network and the position combination network.
  • the main view position detection network is used to determine the position information of the target to be detected in the main view perspective
  • the bird's-eye view position detection network is used to determine the bird's-eye view perspective.
  • the position combination network is used to combine the position information determined respectively by the main view position detection network and the bird's eye position detection network to obtain the spatial position information of the target to be detected.
  • a target detection device includes: an image acquisition unit configured to acquire a two-dimensional image containing a target to be detected; a detection unit configured to input the image to a pre-trained target detection network to obtain the target to be detected Spatial position information, in which the target detection network includes a main view position detection network, a bird's-eye view position detection network and a position combination network.
  • the main view position detection network is used to determine the position information of the target to be detected from the main view perspective
  • the bird's-eye view position detection network is used to To determine the position information of the target to be detected from a bird's-eye view
  • the position combining network is used to combine the position information determined respectively by the main-view position detection network and the bird's-eye position detection network to obtain the spatial position information of the target to be detected.
  • an electronic device includes: one or more processors; a storage device for storing one or more programs; when one or more programs are processed by one or more Execution by a processor causes one or more processors to implement the method described in any of the above implementations.
  • a computer-readable medium is provided with a computer program stored thereon, and when the computer program is executed by a processor, the method described in any of the above implementations is implemented.
  • Figure 1 is an exemplary system architecture diagram in which an embodiment of the present disclosure may be applied;
  • Figure 2 is a flow chart of an embodiment of a target detection method according to the present disclosure
  • Figure 3 is a flow chart of yet another embodiment of a target detection method according to the present disclosure.
  • Figure 4 is a schematic diagram of three-dimensional reconstruction using the target detection network and attitude determination network
  • Figure 5 is a schematic diagram of a three-dimensional anchor point graph in camera coordinate space
  • Figure 6 is a schematic structural diagram of an embodiment of a target detection device according to the present disclosure.
  • FIG. 7 is a schematic structural diagram of an electronic device suitable for implementing embodiments of the present disclosure.
  • FIG. 1 shows an exemplary architecture 100 to which embodiments of the target detection method or target detection device of the present disclosure may be applied.
  • the system architecture 100 may include terminal devices 101, 102, 103, a network 104 and a server 105.
  • the network 104 is a medium used to provide communication links between the terminal devices 101, 102, 103 and the server 105.
  • Network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, among others.
  • the terminal devices 101, 102, 103 interact with the server 105 through the network 104 to receive or send messages, etc.
  • Various client applications can be installed on the terminal devices 101, 102, and 103. For example, browser applications, search applications, image processing applications, 3D modeling applications, etc.
  • the terminal devices 101, 102, and 103 may be hardware or software.
  • the terminal devices 101, 102, and 103 may be various electronic devices, including but not limited to smartphones, tablet computers, e-book readers, laptop computers, desktop computers, and so on.
  • the terminal devices 101, 102, and 103 are software, they can be installed in the electronic devices listed above. It can be implemented as multiple software or software modules (for example, multiple software or software modules used to provide distributed services), or as a single software or software module. There are no specific limitations here.
  • the server 105 may be a server that provides various services, such as a back-end server that provides support for client applications installed on the terminal devices 101, 102, 103.
  • the server can obtain a two-dimensional image containing the target to be detected, and use the target detection network to process the two-dimensional image to obtain the spatial position information of the target to be detected.
  • the two-dimensional image containing the target to be detected can be directly stored locally in the server 105.
  • the server 105 can directly extract the locally stored two-dimensional image containing the target to be detected and process it. At this time, the terminal does not need to be present.
  • the target detection method provided by the embodiment of the present disclosure is generally executed by the server 105, and accordingly, the target detection device is generally provided in the server 105.
  • the terminal devices 101, 102, and 103 can also be installed with image processing applications, and the terminal devices 101, 102, and 103 can also process two-dimensional images containing targets to be detected based on the image processing applications.
  • the target detection method can also be executed by the terminal devices 101, 102, and 103.
  • the target detection device can also be provided in the terminal devices 101, 102, and 103.
  • the server 105 and the network 104 may not be present in the exemplary system architecture 100 .
  • the server 105 may be hardware or software. When the server 105 is hardware, it can be implemented as a distributed server cluster composed of multiple servers or as a single server. When the server 105 is software, it may be implemented as multiple software or software modules (for example, multiple software or software modules used to provide distributed services), or it may be implemented as a single software or software module. There are no specific limitations here.
  • FIG. 2 it shows a process 200 of an embodiment of a target detection method according to the present disclosure.
  • the target detection method comprises the following steps:
  • Step 201 Obtain a two-dimensional image containing the target to be detected.
  • the images may be various types of images, such as pedestrian images, parking lot images, etc.
  • a two-dimensional image can refer to a flat image that does not contain depth information.
  • the target to be detected can be various types of objects, such as people, animals, vehicles, etc.
  • the execution subject the server 105 as shown in Figure 1, etc.
  • the execution subject can obtain the content to be detected from a local or other storage device (the terminal device 101, 102, 103, etc. as shown in Figure 1 or a third party is a data platform, etc.) 2D image of the target.
  • the number of the to-be-detected targets contained in the two-dimensional image may be one or more than two, and may be specifically set according to the actual application scenario.
  • Step 202 Input the image to the pre-trained target detection network to obtain the spatial position information of the target to be detected.
  • the target detection network can be used to detect targets to be detected in the image. Perform detection to determine the spatial location information of the target to be detected.
  • the spatial position information may include plane position information and depth information of the target to be detected, that is, the spatial position information may represent the position of the target to be detected in a three-dimensional space.
  • the target detection network can include a main view position detection network, a bird's eye view position detection network and a position combination network.
  • the main view position detection network can be used to determine the position information of the target to be detected from the main view perspective.
  • the main viewing angle can point to the main viewing angle of the target to be detected.
  • the image of the target to be detected from the main perspective corresponds to the two-dimensional image containing the target to be detected obtained in the above step 201. Therefore, the main view position detection network can determine the plane position information of the target to be detected, and the plane position information can represent the position of the target to be detected in the two-dimensional image.
  • the bird's-eye position detection network can be used to determine the position information of the target to be detected from the bird's-eye view, which can point to the bird's-eye view of the target to be detected.
  • the image of the target to be detected from a bird's-eye view may correspond to the top view of the target to be detected. Therefore, the bird's-eye position detection network can determine the position of the target to be detected in the depth direction, that is, the depth position information, which can represent the depth information of the target to be detected.
  • the position combining network can be used to combine the plane position information determined by the main-view position detection network and the depth position information determined by the bird's-eye position detection network, thereby obtaining the three-dimensional position information of the target to be detected, that is, the spatial position information.
  • the network structures of the main view position detection network, the bird's eye view position detection network and the position combined network can be flexibly constructed or set by technicians according to actual application needs or application scenarios.
  • the main view position detection network and the bird's eye view position detection network can be convolutional neural networks
  • the position combination network can be a regression neural network.
  • the main view position detection network can process the two-dimensional image containing the target to be detected to obtain the position information of the target to be detected from the main view perspective
  • the bird's-eye position detection network can process the two-dimensional image containing the target to be detected to obtain
  • the position combination network combines the position information output by the main-view position detection network and the bird's-eye position detection network to obtain the spatial position information of the target to be detected.
  • the position information output by the main-view position detection network, the position information output by the bird's-eye position detection network, and the spatial position information output by the position combination network can be expressed in various ways according to actual application requirements.
  • the position information output by the main view position detection network can be used to It is represented by the plane position coordinates of each target to be detected.
  • the position information output by the bird's-eye position detection network can be represented by the depth value of each target to be detected.
  • the spatial position information output by the position combination network can be represented by the plane position coordinates of each target to be detected. represented by spatial coordinates composed of depth values.
  • the main view position detection network is used to determine the position information of the target to be detected in the two-dimensional image containing the target to be detected in the main view perspective, thereby obtaining the plane position information of the target to be detected, and at the same time, the bird's-eye view position detection network is used to determine the position information of the target to be detected including the target to be detected.
  • the position information of the target to be detected in the two-dimensional image from a bird's-eye view is used to obtain the depth information of the target to be detected, and then the position combining network is used to combine the plane position information and depth information of the target to be detected to obtain the space of the target to be detected.
  • Position information enables detection of the spatial position of the target to be detected in the two-dimensional image.
  • the target detection network may also include a deviation determination network and a position correction network.
  • the deviation determination network can be used to determine the deviation of the spatial position information
  • the position correction network can be used to correct the spatial position information according to the deviation of the spatial position information output by the deviation determination network, thereby obtaining corrected spatial position information.
  • the network structures of the deviation determination network and the position correction network can be flexibly constructed or set by technicians according to actual application requirements or application scenarios.
  • the bias determination network may be a convolutional neural network and the position correction network may be a regression neural network.
  • the deviation determination network can process the two-dimensional image containing the target to be detected to determine the deviation of the spatial position information, and the position correction network can correct the spatial position information determined by the position combination network according to the deviation of the spatial position information.
  • the deviation of the spatial position information output by the deviation determination network and the corrected spatial position information output by the position correction network can adopt various representation methods according to actual application requirements.
  • the deviation of the spatial position information output by the deviation determination network can be represented by the deviation value of the plane position coordinates and the deviation value of the depth value.
  • the corrected spatial position information output by the position correction network can be represented by the corrected plane position coordinates and the corrected depth value.
  • the depth value is represented by the corrected spatial position information.
  • the output position information has a certain Therefore, using the deviation determination network and the position correction network can alleviate the deviation of the position information determined by the main-view position detection network and the bird's-eye position detection network, thereby improving the accuracy of the corrected spatial position information.
  • the deviation determination network may include a main-view deviation determination network, a bird's-eye deviation determination network and a deviation combination network.
  • the main view deviation determination network can be used to determine the deviation of the position information determined by the main view position detection network
  • the bird's-eye view deviation determination network can be used to determine the deviation of the position information determined by the bird's-eye position detection network
  • the deviation combination network can be used to The deviations determined respectively by the main view deviation determination network and the bird's-eye view deviation determination network are combined to obtain the deviation of the spatial position information determined by the position combination network.
  • the network structures of the main-view deviation determination network, the bird's-eye deviation determination network and the deviation combination network can be flexibly constructed or set by technicians according to actual application needs or application scenarios.
  • the main-view bias determination network and the bird's-eye view bias determination network can be convolutional neural networks
  • the bias combination network can be a regression neural network.
  • the main view deviation determination network can process the two-dimensional image containing the target to be detected to determine the deviation of the position information determined by the main view position detection network
  • the bird's-eye view deviation determination network can process the two-dimensional image containing the target to be detected. Processing is performed to determine the deviation of the position information determined by the bird's-eye view position detection network.
  • the deviation combining network can combine the deviations output by the main-view deviation determination network and the bird's-eye view deviation determination network respectively, thereby obtaining the deviation of the spatial position information.
  • the deviations output by the main-view deviation determination network, the bird's-eye deviation determination network and the deviation combination network can be expressed in various ways according to actual application requirements. For example, it can be represented by a deviation value.
  • the main view deviation determination network and the bird's eye view deviation determination network are used to determine the deviation of the position information output by the main view position detection network and the bird's eye position detection network respectively. Then use the deviation combination network to combine the deviations output by the main view position detection network and the bird's-eye position detection network to obtain the deviation of the spatial position information, which helps to improve the accuracy of determining the deviation of the spatial position information, thereby further improving the deviation based on the spatial position information The accuracy of the corrected spatial position information obtained.
  • the target detection network includes various networks such as main view position detection network, bird's eye view position detection network, position combination network, deviation determination network, position correction network, main view
  • the deviation determination network, the bird's-eye view deviation determination network, and the deviation combination network can be trained using various existing neural network training methods. Each network can be trained individually, or all networks can be combined to achieve end-to-end training.
  • the training data can be obtained from public data sets or pre-set by technicians.
  • the number of targets to be detected included in the two-dimensional image may be at least two. At this time, the spatial position information of multiple targets in the two-dimensional image can be detected.
  • the target detection method proposed in this disclosure can detect the three-dimensional spatial position information of the target to be detected in the two-dimensional image, and the use of the depth information of the target to be detected can help improve the accuracy of target detection and alleviate the problem due to multiple Missed detection caused by interaction between targets or occlusion.
  • the two-dimensional image by processing the two-dimensional image as a whole, the possible impact on the accuracy of the detection results due to the lack of reference information in the multi-stage network-based target detection method can be avoided.
  • the process 300 of the target detection method includes the following steps:
  • Step 301 Obtain a two-dimensional image containing the target to be detected.
  • Step 302 Input the image to the pre-trained target detection network to obtain the spatial position information of the target to be detected.
  • Step 303 Perform three-dimensional reconstruction of the target to be detected based on the spatial position information to obtain a three-dimensional reconstruction result.
  • various existing three-dimensional reconstruction methods can be used to perform three-dimensional reconstruction of the target to be detected using the spatial position information of the target to be detected, and obtain corresponding three-dimensional reconstruction results.
  • the posture information of the target to be detected may be determined first, and then the target to be detected may be located according to the spatial position information and posture information of the target to be detected.
  • the target is reconstructed in three dimensions.
  • the posture information of the target to be detected can be used to describe the posture of the target to be detected in the two-dimensional image.
  • the attitude information of the target to be detected can be expressed in various ways according to actual application requirements.
  • key point coordinates can be used to represent the posture information of the target to be detected.
  • the contour line can be used to represent the posture information of the target to be detected.
  • various methods can be used to determine the attitude information of the target to be detected. For example, you can first determine the category of the target to be detected (such as people, animals, vehicles, etc.), and then use the corresponding attitude detection method to process the two-dimensional image according to the category of the target to be detected to determine the attitude information of the target to be detected. .
  • a pre-trained attitude determination network can be used to determine the attitude information of the target to be detected based on the two-dimensional image containing the target to be detected. For example, a two-dimensional image can be input to the attitude determination network to obtain the attitude information of the target to be detected.
  • the network structure of the attitude determination network can be flexibly constructed or set by technicians according to actual application requirements or application scenarios.
  • the pose determination network may be a convolutional neural network.
  • the attitude determination network can be trained using various existing neural network training methods. Training data can be obtained from public data sets or preset by technical personnel.
  • the attitude determination network can be trained together with the above-mentioned target detection network.
  • the attitude determination network and the above-mentioned target detection network can be regarded as a whole, so that the whole can be trained end-to-end without post-processing. and other additional models to reduce the complexity of network model training, making it easier to implement it in practical applications.
  • the target to be detected After obtaining the attitude information of the target to be detected, the target to be detected can be combined with the attitude information and spatial position information, and various existing three-dimensional reconstruction methods can be used to reconstruct the target to be detected. Using a variety of information can help improve the accuracy of the three-dimensional reconstruction results. accuracy.
  • the target detection network may also include a feature extraction network.
  • the feature extraction network can be used to extract features of the two-dimensional image containing the target to be detected, thereby obtaining the corresponding feature extraction results.
  • the main view position detection network can be used to determine the position information of the target to be detected in the main view perspective based on the feature extraction results
  • the bird's eye position detection network can be used to determine the location information of the target to be detected in the bird's eye perspective based on the feature extraction results.
  • the visual deviation determination network can be used to determine the location information determined by the main visual position detection network based on the feature extraction results.
  • Deviation the bird's-eye view deviation determination network can be used to determine the deviation of the position information determined by the bird's-eye position detection network based on the feature extraction results, and the attitude determination network is used to determine the attitude information of the target to be detected based on the feature extraction results.
  • FIG. 4 shows a schematic diagram of three-dimensional reconstruction using a target detection network and a posture determination network.
  • a two-dimensional image 401 includes a target to be detected (such as seven people in the figure, where two children are located in front of five young people).
  • the two-dimensional image 401 can be first input into a feature extraction network 402 to obtain a feature extraction result (such as a feature map) of the two-dimensional image 401, and then the feature extraction result can be respectively input into a main view position detection network 403, a bird's-eye view position detection network 404, a main view deviation determination network 405, a bird's-eye view deviation determination network 406, and a posture determination network 407.
  • the feature extraction network 402 can use a high-performance deep neural network such as ResNet (Residual Network) and HRNe (High-Resolution Net).
  • ResNet Residual Network
  • HRNe High-Resolution Net
  • the main view position detection network 403, the bird's-eye view position detection network 404, the main view deviation determination network 405, the bird's-eye view deviation determination network 406, and the posture determination network 407 can each include a ResNet Block, etc.
  • the main view position detection network 403 can output the position information of each target to be detected in the main view perspective, which can be represented specifically by a Gaussian heat map as shown in the figure 408 (each circle in the figure represents the main view perspective).
  • the center point of each target to be detected different colors indicate the probability of belonging to the center point).
  • the bird's-eye position detection network 404 can output the position information of each target to be detected in a bird's-eye view, which can be represented by a Gaussian heat map as shown in the figure 409 (each circle in the figure represents the position information of each target to be detected in a bird's-eye perspective). Center point, different colors indicate the probability of belonging to the center point).
  • the bird's-eye view can distinguish the relative front-to-back relationship between the multiple targets to be detected in the imaging area, that is, the relative depth relationship.
  • the main view deviation determination network 405 can output the deviation of the position information 408, which can be represented specifically by a feature map as shown as number 410 in the figure (each arrow in the figure can represent the offset in different directions).
  • the bird's-eye view deviation determination network 406 can output the deviation of the position information 409, which can be represented specifically by the feature map shown as number 411 in the figure (each arrow in the figure represents the offset in the depth direction).
  • the posture determination network 407 can output the feature vector 412 of each target to be detected to represent the posture information of each target to be detected.
  • the position combination network 413 can be used to combine the position information 408 from the main view and the position information 409 from the bird's-eye view to obtain the spatial position information 416 of each target to be detected, which can be represented by a three-dimensional Gaussian heat map (each figure in the figure Each coordinate can describe the probability that the current position is the position of the target to be detected in the three-dimensional space).
  • the deviation combining network 414 can be used to combine the deviation 410 of the position information 408 from the main view perspective and the deviation 411 of the position information 409 from the bird's eye view to obtain the deviation 417 of the spatial position information 416 of each target to be detected (arrows in the figure) represents offset vectors in different directions).
  • the dimension of the position information 408 output by the main view position detection network 403 may be 1*H*W
  • the dimension of the position information 409 output by the bird's-eye view position detection network 404 may be 1*D*W
  • "H", " W" and "D” can represent the dimensions in three directions in the three-dimensional space coordinate system.
  • "H" represents height
  • "W” represents width
  • "D” represents depth.
  • the position combining network 413 can expand the position information 408 along the depth direction (D) and expand the position information 409 along the height direction (H), and then use the expanded position information 408 and 409 After combination, the dimension of the spatial position information 416 is obtained as 1*D*H*W.
  • the dimension of the deviation 410 output by the main-view deviation determination network 405 may be 3*H*W
  • the dimension of the deviation 411 output by the bird's-eye deviation determination network 406 may be 1*D*W.
  • the deviation combining network 414 can expand the deviation 410 along the depth direction (D) and the deviation 411 along the height direction (H), and then combine the expanded deviations 410 and 411 to obtain spatial position information 416
  • the deviation of 417 can be 3*D*H*W.
  • the position correction network 419 can be used to correct the spatial position information 416 of each target to be detected based on the deviation 417 to obtain the corrected spatial position information 420 corresponding to each target to be detected.
  • methods such as maximum value suppression can be used to obtain the corrected spatial position information 420 corresponding to each target to be detected. For example, first set the value of each pixel point lower than the preset threshold to 0, then select the local maximum value, then set the value of the pixel value point near the local maximum value to 0, and set the pixel point where the local maximum value is located. The position is used as the spatial position information of each target to be detected.
  • the feature vector 412 of each target to be detected can be combined with the position information 409 of each target to be detected from a bird's-eye view, that is, the depth information, to obtain a three-dimensional space feature vector 415 of each target to be detected (for example, its dimension can be 128* H*W, at this time, in each Each two-dimensional plane position contains a 128-dimensional feature vector, and the feature vector is aligned at the pixel level with the two-dimensional image containing the target to be detected).
  • the regression method based on the fully connected layer can be used to estimate the posture parameters and shape parameters 418 of each target to be detected, and then use the SMPL (Skinned Multi-Person Linear) model 421 according to The posture parameters and shape parameters 418 of each target to be detected are used to three-dimensionally reconstruct each target to be detected, and a three-dimensional reconstruction result 422 is obtained.
  • SMPL Sed Multi-Person Linear
  • the spatial position information of the target to be detected obtained above is defined in an arbitrary space
  • the spatial position information of the target to be detected needs to be converted into the camera coordinate system (Camera Coordinate System).
  • Camera coordinate system Camera Coordinate System
  • various coordinate system conversion methods can be used for coordinate conversion.
  • the spatial position information in the camera coordinate system can be represented by a three-dimensional anchor map (3D Camera Anchor Map).
  • the corresponding transformation can be implemented using the Normalized Camera Representation (Normalized Camera Representation) based on perspective projection (Weak-Perspective).
  • FIG. 5 shows a schematic diagram of a three-dimensional anchor point graph in camera coordinate space.
  • Anchor points with the same depth can form a depth plane, and multiple depth planes can form a representation space in the depth direction. Therefore, after obtaining the spatial position information 416 and its corresponding deviation 417, the above conversion method can be used to transform the spatial position information 416 and its corresponding deviation 417 into the camera coordinate system respectively to obtain the corresponding anchor point, and then by converting the spatial position
  • the corrected spatial position information is obtained by adding anchor points corresponding to the information 416 and its corresponding deviation 417, that is, the spatial position information of each target to be detected in the camera coordinate space is obtained, and based on this, each object can be obtained. Parameters such as the three-dimensional posture and shape of the target to be detected in the camera coordinate space, as well as the three-dimensional relative position, etc.
  • the end-to-end network formed by combining the target detection network and the attitude determination network can realize the detection of each target to be detected in a single two-dimensional image, the depth relationship between each target to be detected, that is, the relative position, and the three-dimensional reconstruction of each target.
  • Targets to be detected, etc., and waiting for each It has strong robustness in scenarios such as interaction or occlusion between detection targets.
  • the present disclosure provides an embodiment of a target detection device.
  • the device embodiment corresponds to the method embodiment shown in Figure 2.
  • the device can be specifically applied to in various electronic devices.
  • the target detection device 600 includes an image acquisition unit 601 and a detection unit 602 .
  • the image acquisition unit 601 is configured to acquire a two-dimensional image containing the target to be detected;
  • the detection unit 602 is configured to input the image to a pre-trained target detection network to obtain the spatial position information of the target to be detected, where the target detection network It includes the main view position detection network, the bird's-eye view position detection network and the position combination network.
  • the main view position detection network is used to determine the position information of the target to be detected from the main view perspective
  • the bird's-eye view position detection network is used to determine the position of the target to be detected from the bird's-eye view.
  • Information, the position combining network is used to combine the position information determined respectively by the main view position detection network and the bird's eye position detection network to obtain the spatial position information of the target to be detected.
  • the target detection device 600 for the specific processing of the image acquisition unit 601 and the detection unit 602 and the technical effects thereof, please refer to the relevant descriptions of steps 201 and 202 in the corresponding embodiment of Figure 2, respectively. I won’t go into details here.
  • the above-mentioned target detection network also includes a deviation determination network and a position correction network.
  • the deviation determination network is used to determine the deviation of the spatial position information
  • the position correction network is used to determine the deviation of the spatial position information.
  • the spatial position information is corrected to obtain corrected spatial position information.
  • the above-mentioned deviation determination network includes a main-view deviation determination network, a bird's-eye view deviation determination network and a deviation combination network.
  • the main-view deviation determination network is used to determine the position determined by the main-view position detection network.
  • Deviation of information the bird's-eye view deviation determination network is used to determine the deviation of the position information determined by the bird's-eye view position detection network
  • the deviation combination network is used to combine the deviations determined respectively by the principal-view deviation determination network and the bird's-eye view deviation determination network to obtain the deviation of the spatial position information.
  • the above device further includes: a three-dimensional reconstruction unit (not shown in the figure) configured to perform three-dimensional reconstruction of the target to be detected based on spatial position information to obtain a three-dimensional reconstruction result.
  • a three-dimensional reconstruction unit (not shown in the figure) configured to perform three-dimensional reconstruction of the target to be detected based on spatial position information to obtain a three-dimensional reconstruction result.
  • the above device further includes: determining the unit The unit (not shown in the figure) is configured to determine the posture information of the target to be detected; and the above-mentioned three-dimensional reconstruction unit is further configured to perform three-dimensional reconstruction of the target to be detected based on the spatial position information and posture information.
  • the above-mentioned determination unit is further configured to: use a pre-trained posture determination network to determine the posture information of the target to be detected based on the image.
  • the above-mentioned target detection network also includes a feature extraction network, wherein the feature extraction network is used to extract features of the image to obtain feature extraction results; and the main view position detection network is used to extract features based on the features.
  • the extraction result determines the position information of the target to be detected in the main view perspective.
  • the bird's-eye position detection network is used to determine the position information of the target to be detected in the bird's-eye view based on the feature extraction results.
  • the main view deviation determination network is used to determine the main view position based on the feature extraction results.
  • the deviation of the position information determined by the detection network, the bird's-eye view deviation determination network is used to determine the deviation of the position information determined by the bird's-eye view position detection network based on the feature extraction results, and the attitude determination network is used to determine the attitude information of the target to be detected based on the feature extraction results.
  • the number of targets to be detected is at least two.
  • the device acquires a two-dimensional image containing a target to be detected through an image acquisition unit; the detection unit inputs the image into a pre-trained target detection network to obtain the spatial position information of the target to be detected, where the target detection
  • the network includes a main view position detection network, a bird's eye position detection network and a position combination network.
  • the main view position detection network is used to determine the position information of the target to be detected from the main view perspective
  • the bird's eye position detection network is used to determine the location information of the target to be detected from the bird's eye perspective.
  • Position information the position combining network is used to combine the position information determined respectively by the main view position detection network and the bird's eye position detection network to obtain the spatial position information of the target to be detected. Since the depth information of the target to be detected can be learned from the bird's-eye view, the three-dimensional position information of the target to be detected, that is, the spatial position information, can be obtained by combining the position information determined from the main view and the bird's-eye view.
  • FIG. 7 a schematic structural diagram of an electronic device (eg, the server in FIG. 1 ) 700 suitable for implementing embodiments of the present disclosure is shown.
  • Figure 7 shows the server only This is only an example and should not bring any limitations to the functions and scope of use of the embodiments of the present disclosure.
  • the electronic device 700 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 701, which may perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703.
  • a processing device 701 e.g., a central processing unit, a graphics processing unit, etc.
  • RAM random access memory
  • Various programs and data required for the operation of the electronic device 700 are also stored in the RAM 703.
  • the processing device 701, the ROM 702, and the RAM 703 are connected to each other via a bus 704.
  • An input/output (I/O) interface 705 is also connected to the bus 704.
  • the following devices may be connected to the I/O interface 705: input devices 706 including, for example, a touch screen, touch pad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; including, for example, a liquid crystal display (LCD), speakers, vibration An output device 707 such as a computer; a storage device 708 including a magnetic tape, a hard disk, etc.; and a communication device 709. Communication device 709 may allow electronic device 700 to communicate wirelessly or wiredly with other devices to exchange data.
  • FIG. 7 illustrates an electronic device 700 having various means, it should be understood that implementation or availability of all illustrated means is not required. More or fewer means may alternatively be implemented or provided. Each block shown in Figure 7 may represent one device, or may represent multiple devices as needed.
  • embodiments of the present disclosure include a computer program product including a computer program carried on a computer-readable medium, the computer program containing program code for performing the method illustrated in the flowchart.
  • the computer program may be downloaded and installed from the network via communication device 709, or from storage device 708, or from ROM 702.
  • the processing device 701 the above-described functions defined in the method of the embodiment of the present disclosure are performed.
  • the computer-readable medium described in the embodiments of the present disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two.
  • the computer-readable storage medium may be, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: having one or Electrical connection of multiple wires, portable computer disk, hard disk, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), fiber optics, portable compact disk ROM ( CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device.
  • the computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, in which computer-readable program code is carried. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device .
  • Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wire, optical cable, RF (radio frequency), etc., or any suitable combination of the above.
  • the above-mentioned computer-readable medium may be included in the above-mentioned electronic device; it may also exist independently without being assembled into the electronic device.
  • the above-mentioned computer-readable medium carries one or more programs.
  • the electronic device obtains a two-dimensional image containing the target to be detected; inputs the image to the pre-trained target Detection network to obtain the spatial position information of the target to be detected.
  • the target detection network includes the main view position detection network, the bird's eye position detection network and the position combination network.
  • the main view position detection network is used to determine the position of the target to be detected from the main view perspective.
  • the bird's-eye position detection network is used to determine the position information of the target to be detected from a bird's-eye view
  • the position combination network is used to combine the position information determined by the main-view position detection network and the bird's-eye position detection network to obtain the spatial position information of the target to be detected.
  • Computer program code for performing operations of embodiments of the present disclosure may be written in one or more programming languages, including object-oriented programming languages—such as Java, Smalltalk, C++, or a combination thereof, Also included are conventional procedural programming languages—such as the "C" language or similar programming languages.
  • Program code may execute entirely on the user's computer, partially on the user's computer, as a A separate software package executes, partially on the user's computer and partially on a remote computer, or entirely on the remote computer or server.
  • the remote computer can be connected to the user's 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 an Internet service provider). connected via the Internet).
  • LAN local area network
  • WAN wide area network
  • Internet service provider such as an Internet service provider
  • each block in the flowchart or block diagram may represent a module, segment, or portion of code that contains one or more logic functions that implement the specified executable instructions.
  • the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown one after another may actually execute substantially in parallel, or they may sometimes execute in the reverse order, depending on the functionality involved.
  • each block of the block diagram and/or flowchart illustration, and combinations of blocks in the block diagram and/or flowchart illustration can be implemented by special purpose hardware-based systems that perform the specified functions or operations. , or can be implemented using a combination of specialized hardware and computer instructions.
  • the units involved in the embodiments of the present disclosure may be implemented in software or hardware.
  • the described unit may also be provided in a processor, for example, it may be described as: a processor includes an image acquisition unit and a detection unit.
  • the names of these units do not constitute a limitation on the unit itself under certain circumstances.
  • the image acquisition unit can also be described as "a unit that acquires a two-dimensional image containing a target to be detected.”

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Geometry (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Remote Sensing (AREA)
  • Computer Graphics (AREA)
  • Image Analysis (AREA)

Abstract

本公开的实施例公开了目标检测方法和装置。该方法的一具体实施方式包括：获取包含待检测目标的二维图像；将图像输入至预先训练的目标检测网络，得到待检测目标的空间位置信息，其中，目标检测网络包括主视位置检测网络、鸟瞰位置检测网络和位置结合网络，主视位置检测网络用于确定主视视角下待检测目标的位置信息，鸟瞰位置检测网络用于确定鸟瞰视角下待检测目标的位置信息，位置结合网络用于结合主视位置检测网络和鸟瞰位置检测网络分别确定的位置信息以得到待检测目标的空间位置信息。

Description

目标检测方法和装置
相关申请的交叉引用
本专利申请要求于2022年9月19日提交的、申请号为202211136508.5、发明名称为“目标检测方法和装置”的中国专利申请的优先权，该申请的全文以引用的方式并入本申请中。
技术领域
本公开的实施例涉及计算机技术领域，具体涉及目标检测方法和装置。
背景技术
目标检测是计算机视觉领域的一个重要研究方向，其在智能交通、智能辅助驾驶和视频监控等领域都具有良好的应用前景。随着深度学习的快速发展，利用深度学习实现目标检测是重点研究的目标检测算法之一。
发明内容
本公开的实施例提出了目标检测方法和装置。
在一个或多个实施例中，提供了一种目标检测方法，该方法包括：获取包含待检测目标的二维图像；将图像输入至预先训练的目标检测网络，得到待检测目标的空间位置信息，其中，目标检测网络包括主视位置检测网络、鸟瞰位置检测网络和位置结合网络，主视位置检测网络用于确定主视视角下待检测目标的位置信息，鸟瞰位置检测网络用于确定鸟瞰视角下待检测目标的位置信息，位置结合网络用于结合主视位置检测网络和鸟瞰位置检测网络分别确定的位置信息以得到待检测目标的空间位置信息。
在一个或多个实施例中，提供了一种目标检测装置，该装置包括：图像获取单元，被配置成获取包含待检测目标的二维图像；检测单元，被配置成将图像输入至预先训练的目标检测网络，得到待检测目标的 空间位置信息，其中，目标检测网络包括主视位置检测网络、鸟瞰位置检测网络和位置结合网络，主视位置检测网络用于确定主视视角下待检测目标的位置信息，鸟瞰位置检测网络用于确定鸟瞰视角下待检测目标的位置信息，位置结合网络用于结合主视位置检测网络和鸟瞰位置检测网络分别确定的位置信息以得到待检测目标的空间位置信息。
在一个或多个实施例中，提供了一种电子设备，该电子设备包括：一个或多个处理器；存储装置，用于存储一个或多个程序；当一个或多个程序被一个或多个处理器执行，使得一个或多个处理器实现如上任一实现方式描述的方法。
在一个或多个实施例中，提供了一种计算机可读介质，其上存储有计算机程序，该计算机程序被处理器执行时实现如上任一实现方式描述的方法。
附图说明
通过阅读参照以下附图所作的对非限制性实施例所作的详细描述，本公开的其它特征、目的和优点将会变得更明显：
图1是本公开的一个实施例可以应用于其中的示例性系统架构图；
图2是根据本公开的目标检测方法的一个实施例的流程图；
图3是根据本公开的目标检测方法的又一个实施例的流程图；
图4是利用目标检测网络和姿态确定网络进行三维重建的示意图；
图5是相机坐标空间下的三维锚点图的示意图；
图6是根据本公开的目标检测装置的一个实施例的结构示意图；
图7是适于用来实现本公开的实施例的电子设备的结构示意图。
具体实施方式
下面结合附图和实施例对本公开作进一步的详细说明。可以理解的是，此处所描述的具体实施例仅仅用于解释相关发明，而非对该发 明的限定。另外还需要说明的是，为了便于描述，附图中仅示出了与有关发明相关的部分。
需要说明的是，在不冲突的情况下，本公开中的实施例及实施例中的特征可以相互组合。下面将参考附图并结合实施例来详细说明本公开。
图1示出了可以应用本公开的目标检测方法或目标检测装置的实施例的示例性架构100。
如图1所示，系统架构100可以包括终端设备101、102、103，网络104和服务器105。网络104用以在终端设备101、102、103和服务器105之间提供通信链路的介质。网络104可以包括各种连接类型，例如有线、无线通信链路或者光纤电缆等等。
终端设备101、102、103通过网络104与服务器105交互，以接收或发送消息等。终端设备101、102、103上可以安装有各种客户端应用。例如浏览器类应用、搜索类应用、图像处理类应用、三维建模类应用等等。
终端设备101、102、103可以是硬件，也可以是软件。当终端设备101、102、103为硬件时，可以是各种电子设备，包括但不限于智能手机、平板电脑、电子书阅读器、膝上型便携计算机和台式计算机等等。当终端设备101、102、103为软件时，可以安装在上述所列举的电子设备中。其可以实现成多个软件或软件模块(例如用来提供分布式服务的多个软件或软件模块)，也可以实现成单个软件或软件模块。在此不做具体限定。
服务器105可以是提供各种服务的服务器，例如为终端设备101、102、103上安装的客户端应用提供支持的后端服务器。服务器可以获取包含待检测目标的二维图像，并利用目标检测网络对二维图像进行处理得到待检测目标的空间位置信息。
需要说明的是，包含待检测目标的二维图像可以直接存储在服务器105的本地，服务器105可以直接提取本地所存储的包含待检测目标的二维图像并进行处理，此时，可以不存在终端设备101、102、103和网络104)。
需要说明的是，本公开的实施例所提供的目标检测方法一般由服务器105执行，相应地，目标检测装置一般设置于服务器105中。
还需要指出的是，终端设备101、102、103中也可以安装有图像处理类应用，终端设备101、102、103也可以基于图像处理类应用对包含待检测目标的二维图像进行处理，此时，目标检测方法也可以由终端设备101、102、103执行，相应地，目标检测装置也可以设置于终端设备101、102、103中。此时，示例性系统架构100可以不存在服务器105和网络104。
需要说明的是，服务器105可以是硬件，也可以是软件。当服务器105为硬件时，可以实现成多个服务器组成的分布式服务器集群，也可以实现成单个服务器。当服务器105为软件时，可以实现成多个软件或软件模块(例如用来提供分布式服务的多个软件或软件模块)，也可以实现成单个软件或软件模块。在此不做具体限定。
应该理解，图1中的终端设备、网络和服务器的数目仅仅是示意性的。根据实现需要，可以具有任意数目的终端设备、网络和服务器。
继续参考图2，其示出了根据本公开的目标检测方法的一个实施例的流程200。该目标检测方法包括以下步骤：
步骤201，获取包含待检测目标的二维图像。
在本实施例中，图像可以是各种类型的图像，如行人图像、停车场图像等等。二维图像可以指不包含深度信息的平面图像。待检测目标可以是各种类型的对象，如人、动物、车辆等等。具体地，执行主体(如图1所示的服务器105等)可以从本地或其他存储设备(如图1所示的终端设备101、102、103等或第三方是数据平台等)获取包含待检测目标的二维图像。
需要说明的是，二维图像中包含的待检测目标的数量可以单个，也可以为两个以上，具体可以根据实际的应用场景设置。
步骤202，将图像输入至预先训练的目标检测网络，得到待检测目标的空间位置信息。
在本实施例中，目标检测网络可以用于对图像中的待检测目标进 行检测以确定待检测目标的空间位置信息。其中，空间位置信息可以包括待检测目标的平面位置信息和深度信息，即空间位置信息可以表示待检测目标在三维空间中的位置。
目标检测网络可以包括主视位置检测网络、鸟瞰位置检测网络和位置结合网络。其中，主视位置检测网络可以用于确定主视视角下待检测目标的位置信息。主视视角可以指针对待检测目标的主视视角。一般地，主视视角下待检测目标的图像对应于上述步骤201获取的包含待检测目标的二维图像。因此，主视位置检测网络可以确定待检测目标的平面位置信息，该平面位置信息可以表示待检测目标在二维图像中的位置。
鸟瞰位置检测网络可以用于确定鸟瞰视角下待检测目标的位置信息，鸟瞰视角可以指针对待检测目标的俯视视角。一般地，鸟瞰视角下待检测目标的图像可以对应于待检测目标的俯视图。因此，鸟瞰位置检测网络可以确定待检测目标在深度方向的位置，即深度位置信息，该深度位置信息可以表示待检测目标的深度信息。
位置结合网络可以用于结合主视位置检测网络确定的平面位置信息和鸟瞰位置检测网络确定的深度位置信息，从而得到待检测目标的三维位置信息即空间位置信息。
其中，主视位置检测网络、鸟瞰位置检测网络和位置结合网络的网络结构可以由技术人员根据实际的应用需求或应用场景灵活构建或设置。例如，主视位置检测网络和鸟瞰位置检测网络可以是卷积神经网络，位置结合网络可以是回归神经网络。具体地，主视位置检测网络可以对包含待检测目标的二维图像进行处理得到待检测目标在主视视角下的位置信息，鸟瞰位置检测网络可以对包含待检测目标的二维图像进行处理得到待检测目标在鸟瞰视角下的位置信息，位置结合网络对主视位置检测网络和鸟瞰位置检测网络分别输出的位置信息进行结合得到待检测目标的空间位置信息。
主视位置检测网络输出的位置信息、鸟瞰位置检测网络输出的位置信息和位置结合网络输出的空间位置信息可以根据实际的应用需求采用各种表示方法。例如，主视位置检测网络输出的位置信息可以利 用各待检测目标的平面位置坐标来表示，鸟瞰位置检测网络输出的位置信息可以利用各待检测目标的深度值来表示，位置结合网络输出的空间位置信息可以利用各待检测目标的平面位置坐标和深度值组成的空间坐标来表示。
利用主视位置检测网络确定包含待检测目标的二维图像中的待检测目标在主视视角下的位置信息，从而得到待检测目标的平面位置信息，同时利用鸟瞰位置检测网络确定包含待检测目标的二维图像中的待检测目标在鸟瞰视角下的位置信息，从而得到待检测目标的深度信息，再利用位置结合网络结合待检测目标的平面位置信息和深度信息，从而得到待检测目标的空间位置信息，实现对二维图像中的待检测目标的空间位置的检测。
在本实施例的一些可选的实现方式中，目标检测网络还可以包括偏差确定网络和位置校正网络。其中，偏差确定网络可以用于确定空间位置信息的偏差，位置校正网络可以用于根据偏差确定网络输出的空间位置信息的偏差对空间位置信息进行校正，从而得到校正后的空间位置信息。
其中，偏差确定网络和位置校正网络的网络结构可以由技术人员根据实际的应用需求或应用场景灵活构建或设置。例如，偏差确定网络可以是卷积神经网络，位置校正网络可以是回归神经网络。具体地，偏差确定网络可以对包含待检测目标的二维图像进行处理以确定空间位置信息的偏差，位置校正网络可以根据空间位置信息的偏差对位置结合网络确定的空间位置信息进行校正。
偏差确定网络输出的空间位置信息的偏差和位置校正网络输出的校正后的空间位置信息可以根据实际的应用需求采用各种表示方法。例如，偏差确定网络输出的空间位置信息的偏差可以利用平面位置坐标的偏差值和深度值的偏差值来表示，位置校正网络输出的校正后空间位置信息可以利用校正后的平面位置坐标和校正后的深度值组成的校正后的空间位置信息来表示。
由于主视位置检测网络和鸟瞰位置检测网络等在对二维图像的处理过程中通常会有离散化等步骤，从而导致输出的位置信息具有一定 偏差，因此，利用偏差确定网络和位置校正网络可以缓解主视位置检测网络和鸟瞰位置检测网络等确定的位置信息的偏差，从而提升校正后的空间位置信息的准确性。
可选地，偏差确定网络可以包括主视偏差确定网络、鸟瞰偏差确定网络和偏差结合网络。其中，主视偏差确定网络可以用于确定主视位置检测网络所确定的位置信息的偏差，鸟瞰偏差确定网络可以用于确定鸟瞰位置检测网络所确定的位置信息的偏差，偏差结合网络可以用于结合主视偏差确定网络和鸟瞰偏差确定网络分别确定的偏差，以得到位置结合网络确定的空间位置信息的偏差。
其中，主视偏差确定网络、鸟瞰偏差确定网络和偏差结合网络的网络结构可以由技术人员根据实际的应用需求或应用场景灵活构建或设置。例如，主视偏差确定网络和鸟瞰偏差确定网络可以是卷积神经网络，偏差结合网络可以是回归神经网络。具体地，主视偏差确定网络可以对包含待检测目标的二维图像进行处理以确定主视位置检测网络所确定的位置信息的偏差，鸟瞰偏差确定网络可以对包含待检测目标的二维图像进行处理以确定鸟瞰位置检测网络所确定的位置信息的偏差，偏差结合网络可以对主视偏差确定网络和鸟瞰偏差确定网络分别输出的偏差进行结合，从而得到空间位置信息的偏差。
主视偏差确定网络、鸟瞰偏差确定网络和偏差结合网络分别输出的偏差可以根据实际的应用需求采用各种表示方法。例如，可以采用偏差值进行表示。
由于主视位置检测网络和鸟瞰位置检测网络的处理过程是相对独立的，因此利用主视偏差确定网络和鸟瞰偏差确定网络分别确定主视位置检测网络和鸟瞰位置检测网络输出的位置信息的偏差，再利用偏差结合网络结合主视位置检测网络和鸟瞰位置检测网络分别输出的偏差得到空间位置信息的偏差，有助于提升确定空间位置信息的偏差的精确性，从而进一步提升基于空间位置信息的偏差得到的校正后的空间位置信息的准确性。
上述目标检测网络所包括的各个网络如主视位置检测网络、鸟瞰位置检测网络、位置结合网络、偏差确定网络、位置校正网络、主视 偏差确定网络、鸟瞰偏差确定网络和偏差结合网络等可以采用现有的各种神经网络的训练方法进行训练得到。其中的每个网络可以单独训练，也可以将所有的网络联合实现端到端的训练，训练数据可以从公开数据集获取，也可以由技术人员预先设置。
在本实施例的一些可选的实现方式中，二维图像中包含的待检测目标的数量可以为至少两个。此时，可以实现对二维图像中的多目标的空间位置信息的检测。
现有的基于单阶段网络的目标检测方法通常没有对各目标的深度信息进行估计，由于对深度方向上的忽略，容易在多目标之间的交互和/或遮挡等情况下出现漏检等问题，而基于多阶段网络的目标检测方法由于后续阶段处理的图像不是最初完整的包含多个目标的图像，从而会导致缺少一些参照信息，影响确定的深度信息的准确性。针对这些问题，本公开提出的目标检测方法可以实现对二维图像中的待检测目标的三维空间位置信息的检测，而且利用待检测目标的深度信息有助于提升目标检测准确性，缓解由于多目标之间的交互或遮挡导致的漏检等情况。同时，通过对二维图像整体进行处理，可以避免基于多阶段网络的目标检测方法可能出现的由于缺乏参照信息导致对检测结果准确度的影响。
进一步参考图3，其示出了目标检测方法的又一个实施例的流程300。该目标检测方法的流程300，包括以下步骤：
步骤301，获取包含待检测目标的二维图像。
步骤302，将图像输入至预先训练的目标检测网络，得到待检测目标的空间位置信息。
步骤303，根据空间位置信息对待检测目标进行三维重建，得到三维重建结果。
在本实施例中，在得到待检测目标的空间位置信息之后，可以利用现有的各种三维重建方法利用待检测目标的空间位置信息对待检测目标进行三维重建，得到对应的三维重建结果。
在本实施例的一些可选的实现方式中，可以先确定待检测目标的姿态信息，然后根据待检测目标的空间位置信息和姿态信息，对待检 测目标进行三维重建。
其中，待检测目标的姿态信息可以用于描述二维图像中待检测目标的姿态。待检测目标的姿态信息可以根据实际的应用需求采用各种表示方法。例如，可以利用关键点坐标来表示待检测目标的姿态信息。又例如，可以利用轮廓线来表示待检测目标的姿态信息。
具体地，可以采用各种方法确定待检测目标的姿态信息。例如，可以先确定待检测目标的类别(如人、动物、车辆等等)，然后根据待检测目标的类别，采用对应的姿态检测方法对二维图像进行处理，以确定待检测目标的姿态信息。
可选地，可以利用预先训练的姿态确定网络，根据包含待检测目标的二维图像确定待检测目标的姿态信息。例如，可以将二维图像输入至姿态确定网络，得到待检测目标的姿态信息。其中，姿态确定网络的网络结构可以由技术人员根据实际的应用需求或应用场景灵活构建或设置。例如，姿态确定网络可以是卷积神经网络。
姿态确定网络可以利用现有的各种神经网络的训练方法进行训练得到。训练数据可以从公开数据集获取，也可以由技术人员预先设置。
可选地，姿态确定网络可以和上述目标检测网络一起训练，此时，可以将姿态确定网络和上述目标检测网络视为一个整体，从而可以对这一整体进行端到端的训练，不需要后处理等额外模型，降低网络模型训练的复杂度，从而易于落地于实际应用中。
在得到待检测目标的姿态信息之后，可以结合待检测目标的姿态信息和空间位置信息，利用现有的各种三维重建方法对待检测目标进行重建，利用多种信息有助于提升三维重建结果的准确性。
在本实施例的一些可选的实现方式中，目标检测网络还可以包括特征提取网络。其中，特征提取网络可以用于提取包含待检测目标的二维图像的特征，从而得到对应的特征提取结果。
此时，主视位置检测网络可以用于根据特征提取结果确定主视视角下待检测目标的位置信息，鸟瞰位置检测网络可以用于根据特征提取结果确定鸟瞰视角下待检测目标的位置信息，主视偏差确定网络可以用于根据特征提取结果确定主视位置检测网络所确定的位置信息的 偏差，鸟瞰偏差确定网络可以用于根据特征提取结果确定鸟瞰位置检测网络所确定的位置信息的偏差，姿态确定网络用于根据特征提取结果确定待检测目标的姿态信息。
作为示例，继续参见图4，其示出了利用目标检测网络和姿态确定网络进行三维重建的示意图。如图4所示，二维图像401包括待检测目标(如图中的七个人，其中，两个孩童位于五个青年的前面)。具体地，可以先将二维图像401输入至特征提取网络402，得到二维图像401的特征提取结果(如特征图)，然后可以将特征提取结果分别输入主视位置检测网络403、、鸟瞰位置检测网络404、主视偏差确定网络405、鸟瞰偏差确定网络406和姿态确定网络407。例如，特征提取网络402可以采用如ResNet(Residual Network)、HRNe(High-Resolution Net)等较高性能的深度神经网络。此时，主视位置检测网络403、、鸟瞰位置检测网络404、主视偏差确定网络405、鸟瞰偏差确定网络406和姿态确定网络407可以均包括一个ResNet Block等。
其中，主视位置检测网络403可以输出各个待检测目标在主视视角下的位置信息，具体可以采用如图中编号408所示的高斯热图来表示(图中的各个圆表示主视视角下各个待检测目标的中心点，不同颜色表示属于中心点的概率)。鸟瞰位置检测网络404可以输出各个待检测目标在鸟瞰视角下的位置信息，具体可以采用如图中编号409所示的高斯热图来表示(图中的各个圆表示鸟瞰视角下各个待检测目标的中心点，不同颜色表示属于中心点的概率)。在如二维图像401所示的包含多个待检测目标的情况下，鸟瞰视角可以区分成像区域内多个待检测目标之间的相对前后关系，即相对深度关系。主视偏差确定网络405可以输出位置信息408的偏差，具体可以采用如图中编号410所示的特征图来表示(图中的各个箭头可以表示不同方向的偏移量)。鸟瞰偏差确定网络406可以输出位置信息409的偏差，具体可以采用如图中编号411所示的特征图来表示(图中的各箭头表示深度方向的偏移量)。姿态确定网络407可以输出各个待检测目标的特征向量412来表示各个待检测目标的姿态信息。
然后，可以利用位置结合网络413结合主视视角下的位置信息408和鸟瞰视角下的位置信息409得到各待检测目标的空间位置信息416，可以采用三维的高斯热图来表示(图中的每个坐标都可以描述当前位置是待检测目标在三维空间中的位置的概率)。同时，可以利用偏差结合网络414结合主视视角下的位置信息408的偏差410和鸟瞰视角下的位置信息409的偏差411得到各待检测目标的空间位置信息416的偏差417(图中的各箭头表示在不同方向的偏移向量)。
例如，主视位置检测网络403输出的位置信息408的维度可以为1*H*W，鸟瞰位置检测网络404输出的位置信息409的维度可以为1*D*W，其中，“H”、“W”和“D”可以表示三维空间坐标系中三个方向上的尺寸。如“H”表示高度、“W”表示宽度，以及“D”表示深度。此时，位置结合网络413可以将位置信息408沿着深度方向(D)进行扩张(expand)，同时将位置信息409沿着高度方向(H)进行扩张，然后将扩张后的位置信息408和409进行组合，得到空间位置信息416的维度即可以为1*D*H*W。主视偏差确定网络405输出的偏差410的维度可以为3*H*W，鸟瞰偏差确定网络406输出的偏差411的维度可以为1*D*W。此时，偏差结合网络414可以将偏差410沿着深度方向(D)扩张，同时将偏差411沿着高度方向(H)扩张，然后将扩展后的偏差410和411进行组合，得到空间位置信息416的偏差417的偏差可以为3*D*H*W。
然后，可以利用位置校正网络419根据偏差417对各待检测目标的空间位置信息416进行校正，得到各待检测目标对应的校正后的空间位置信息420。具体地，可以利用极大值抑制等方法得到各待检测目标对应的校正后的空间位置信息420。例如，先将各个低于预设阈值的像素点的取值置0，然后选取局部最大值，再将局部最大值附近的像素值点的取值置0，并将局部最大值所在的像素点的位置作为各待检测目标的空间位置信息。
另外，可以在各待检测目标的特征向量412上结合各待检测目标的鸟瞰视角下的位置信息409，即深度信息，得到各个待检测目标的三维空间特征向量415(例如其维度可以为128*H*W，此时，在每个 二维平面位置上都包含一个128维的特征向量，而该特征向量在像素级别上与包含待检测目标的二维图像对齐)。然后可以根据各个待检测目标的三维空间特征向量415，利用基于全连接层的回归方法等估计各待检测目标的姿态参数和形状参数418等，再利用SMPL(Skinned Multi-Person Linear)模型421根据各待检测目标的姿态参数和形状参数418等对各待检测目标进行三维重建，得到三维重建结果422。
此外，需要说明的是，由于上述得到的待检测目标的空间位置信息是在任意空间中定义的，因此在需要将待检测目标在真实成像空间中进行三维重建时，需要将得到的待检测目标的空间位置信息转换到相机坐标系(Camera Coordinate System)。具体地，可以采用各种坐标系转换方法进行坐标转换。
例如，可以将相机坐标系下的空间位置信息利用三维锚点图(3D Camera Anchor Map)进行表示。此时，可以利用基于透视投影Weak-Perspective)的归一化相机表征(Normalized Camera Representation)实现对应变换。具体地，可以定义归一化因子si＝(ditan(FOV/2))-1，并通过变换得到和其中，(xi、yi、di)可以得到的空间位置信息、“tan”表示正切函数、“FOV”表示视场角，以及表示对应转换后的锚点值。如图5所示，其示出了相机坐标空间下的三维锚点图的示意图。具有相同深度的锚点可以组成一个深度平面，多个深度平面即可以构成深度方向的表征空间。因此，在得到空间位置信息416和其对应的偏差417之后，可以利用上述转换方法将空间位置信息416和其对应的偏差417分别进行相机坐标系转换，得到对应的锚点，再通过将空间位置信息416和其对应的偏差417分别对应的锚点相加等结合方式得到校正后的空间位置信息，即得到每个待检测目标在相机坐标空间下的空间位置信息，进而基于此也可以得到各待检测目标在相机坐标空间下的三维姿态和形状等参数，以及三维相对位置等等。
由此，结合目标检测网络和姿态确定网络等形成的端到端的网络可以实现对单张二维图像中的各待检测目标的检测、各待检测目标之间的深度关系即相对位置，以及三维重建各待检测目标等，且在各待 检测目标之间具有交互或遮挡等场景下具有较强的鲁棒性。
进一步参考图6，作为对上述各图所示方法的实现，本公开提供了目标检测装置的一个实施例，该装置实施例与图2所示的方法实施例相对应，该装置具体可以应用于各种电子设备中。
如图6所示，本实施例提供的目标检测装置600包括图像获取单元601和检测单元602。其中，图像获取单元601被配置成获取包含待检测目标的二维图像；检测单元602被配置成将图像输入至预先训练的目标检测网络，得到待检测目标的空间位置信息，其中，目标检测网络包括主视位置检测网络、鸟瞰位置检测网络和位置结合网络，主视位置检测网络用于确定主视视角下待检测目标的位置信息，鸟瞰位置检测网络用于确定鸟瞰视角下待检测目标的位置信息，位置结合网络用于结合主视位置检测网络和鸟瞰位置检测网络分别确定的位置信息以得到待检测目标的空间位置信息。
在本实施例中，目标检测装置600中：图像获取单元601和检测单元602的具体处理及其所带来的技术效果可分别参考图2对应实施例中的步骤201和步骤202的相关说明，在此不再赘述。
在本实施例的一些可选的实现方式中，上述目标检测网络还包括偏差确定网络和位置校正网络，偏差确定网络用于确定空间位置信息的偏差，位置校正网络用于根据空间位置信息的偏差对空间位置信息进行校正，得到校正后的空间位置信息。
在本实施例的一些可选的实现方式中，上述偏差确定网络包括主视偏差确定网络、鸟瞰偏差确定网络和偏差结合网络，主视偏差确定网络用于确定主视位置检测网络所确定的位置信息的偏差，鸟瞰偏差确定网络用于确定鸟瞰位置检测网络所确定的位置信息的偏差，偏差结合网络用于结合主视偏差确定网络和鸟瞰偏差确定网络分别确定的偏差以得到空间位置信息的偏差。
在本实施例的一些可选的实现方式中，上述装置还包括：三维重建单元(图中未示出)被配置成根据空间位置信息对待检测目标进行三维重建，得到三维重建结果。
在本实施例的一些可选的实现方式中，上述装置还包括：确定单 元(图中未示出)被配置成确定待检测目标的姿态信息；以及上述三维重建单元进一步被配置成根据空间位置信息和姿态信息，对待检测目标进行三维重建。
在本实施例的一些可选的实现方式中，上述确定单元进一步被配置成：利用预先训练的姿态确定网络根据图像确定待检测目标的姿态信息。
在本实施例的一些可选的实现方式中，上述目标检测网络还包括特征提取网络，其中，特征提取网络用于提取图像的特征，得到特征提取结果；以及主视位置检测网络用于根据特征提取结果确定主视视角下待检测目标的位置信息，鸟瞰位置检测网络用于根据特征提取结果确定鸟瞰视角下待检测目标的位置信息，主视偏差确定网络用于根据特征提取结果确定主视位置检测网络所确定的位置信息的偏差，鸟瞰偏差确定网络用于根据特征提取结果确定鸟瞰位置检测网络所确定的位置信息的偏差，姿态确定网络用于根据特征提取结果确定待检测目标的姿态信息。
在本实施例的一些可选的实现方式中，待检测目标的数量为至少两个。
本公开的上述实施例提供的装置，通过图像获取单元获取包含待检测目标的二维图像；检测单元将图像输入至预先训练的目标检测网络，得到待检测目标的空间位置信息，其中，目标检测网络包括主视位置检测网络、鸟瞰位置检测网络和位置结合网络，主视位置检测网络用于确定主视视角下待检测目标的位置信息，鸟瞰位置检测网络用于确定鸟瞰视角下待检测目标的位置信息，位置结合网络用于结合主视位置检测网络和鸟瞰位置检测网络分别确定的位置信息以得到待检测目标的空间位置信息。由于鸟瞰视角下可以学习待检测目标的深度信息，从而结合主视视角下和鸟瞰视角下确定的位置信息可以得到待检测目标的三维位置信息即空间位置信息。
下面参考图7，其示出了适于用来实现本公开的实施例的电子设备(例如图1中的服务器)700的结构示意图。图7示出的服务器仅 仅是一个示例，不应对本公开的实施例的功能和使用范围带来任何限制。
如图7所示，电子设备700可以包括处理装置(例如中央处理器、图形处理器等)701，其可以根据存储在只读存储器(ROM)702中的程序或者从存储装置708加载到随机访问存储器(RAM)703中的程序而执行各种适当的动作和处理。在RAM 703中，还存储有电子设备700操作所需的各种程序和数据。处理装置701、ROM 702以及RAM 703通过总线704彼此相连。输入/输出(I/O)接口705也连接至总线704。
通常，以下装置可以连接至I/O接口705：包括例如触摸屏、触摸板、键盘、鼠标、摄像头、麦克风、加速度计、陀螺仪等的输入装置706；包括例如液晶显示器(LCD)、扬声器、振动器等的输出装置707；包括例如磁带、硬盘等的存储装置708；以及通信装置709。通信装置709可以允许电子设备700与其他设备进行无线或有线通信以交换数据。虽然图7示出了具有各种装置的电子设备700，但是应理解的是，并不要求实施或具备所有示出的装置。可以替代地实施或具备更多或更少的装置。图7中示出的每个方框可以代表一个装置，也可以根据需要代表多个装置。
特别地，根据本公开的实施例，上文参考流程图描述的过程可以被实现为计算机软件程序。例如，本公开的实施例包括一种计算机程序产品，其包括承载在计算机可读介质上的计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中，该计算机程序可以通过通信装置709从网络上被下载和安装，或者从存储装置708被安装，或者从ROM 702被安装。在该计算机程序被处理装置701执行时，执行本公开的实施例的方法中限定的上述功能。
需要说明的是，本公开的实施例所述的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于：具有一个或 多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本公开的实施例中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本公开的实施例中，计算机可读信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读信号介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：电线、光缆、RF(射频)等等，或者上述的任意合适的组合。
上述计算机可读介质可以是上述电子设备中所包含的；也可以是单独存在，而未装配入该电子设备中。上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被该电子设备执行时，使得该电子设备：获取包含待检测目标的二维图像；将图像输入至预先训练的目标检测网络，得到待检测目标的空间位置信息，其中，目标检测网络包括主视位置检测网络、鸟瞰位置检测网络和位置结合网络，主视位置检测网络用于确定主视视角下待检测目标的位置信息，鸟瞰位置检测网络用于确定鸟瞰视角下待检测目标的位置信息，位置结合网络用于结合主视位置检测网络和鸟瞰位置检测网络分别确定的位置信息以得到待检测目标的空间位置信息。
可以以一种或多种程序设计语言或其组合来编写用于执行本公开的实施例的操作的计算机程序代码，所述程序设计语言包括面向对象的程序设计语言—诸如Java、Smalltalk、C++，还包括常规的过程式程序设计语言—诸如“C”语言或类似的程序设计语言。程序代码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)。
附图中的流程图和框图，图示了按照本公开各种实施例的系统、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、或代码的一部分，该模块、程序段、或代码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。也应当注意，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，两个接连地表示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操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。
描述于本公开的实施例中所涉及到的单元可以通过软件的方式实现，也可以通过硬件的方式来实现。所描述的单元也可以设置在处理器中，例如，可以描述为：一种处理器包括图像获取单元和检测单元。其中，这些单元的名称在某种情况下并不构成对该单元本身的限定，例如，图像获取单元还可以被描述为“获取包含待检测目标的二维图像的单元”。
以上描述仅为本公开的较佳实施例以及对所运用技术原理的说明。本领域技术人员应当理解，本公开的实施例中所涉及的公开范围，并不限于上述技术特征的特定组合而成的技术方案，同时也应涵盖在不脱离本公开的构思的情况下，由上述技术特征或其等同特征进行任意组合而形成的其它技术方案。例如上述特征与本公开的实施例中公开的(但不限于)具有类似功能的技术特征进行互相替换而形成的技术方案。

Claims (18)

  1. 一种目标检测方法，包括：
    获取包含待检测目标的二维图像；
    将所述图像输入至预先训练的目标检测网络，得到所述待检测目标的空间位置信息，其中，所述目标检测网络包括主视位置检测网络、鸟瞰位置检测网络和位置结合网络，所述主视位置检测网络用于确定主视视角下所述待检测目标的位置信息，所述鸟瞰位置检测网络用于确定鸟瞰视角下所述待检测目标的位置信息，所述位置结合网络用于结合所述主视位置检测网络和鸟瞰位置检测网络分别确定的位置信息以得到所述待检测目标的空间位置信息。
  2. 根据权利要求1所述的方法，其中，所述目标检测网络还包括偏差确定网络和位置校正网络，所述偏差确定网络用于确定所述空间位置信息的偏差，所述位置校正网络用于根据所述空间位置信息的偏差对所述空间位置信息进行校正，得到校正后的空间位置信息。
  3. 根据权利要求2所述的方法，其中，所述偏差确定网络包括主视偏差确定网络、鸟瞰偏差确定网络和偏差结合网络，所述主视偏差确定网络用于确定所述主视位置检测网络所确定的位置信息的偏差，所述鸟瞰偏差确定网络用于确定所述鸟瞰位置检测网络所确定的位置信息的偏差，所述偏差结合网络用于结合所述主视偏差确定网络和鸟瞰偏差确定网络分别确定的偏差以得到所述空间位置信息的偏差。
  4. 根据权利要求3所述的方法，其中，所述方法还包括：
    根据所述空间位置信息对所述待检测目标进行三维重建，得到三维重建结果。
  5. 根据权利要求4所述的方法，其中，所述方法还包括：
    确定所述待检测目标的姿态信息；以及
    所述根据所述空间位置信息对所述待检测目标进行三维重建，包括：
    根据所述空间位置信息和姿态信息，对所述待检测目标进行三维重建。
  6. 根据权利要求5所述的方法，其中，所述确定所述待检测目标的姿态信息，包括：
    利用预先训练的姿态确定网络根据所述图像确定所述待检测目标的姿态信息。
  7. 根据权利要求6所述的方法，其中，所述目标检测网络还包括特征提取网络，其中，所述特征提取网络用于提取所述图像的特征，得到特征提取结果；以及
    所述主视位置检测网络用于根据所述特征提取结果确定主视视角下所述待检测目标的位置信息，所述鸟瞰位置检测网络用于根据所述特征提取结果确定鸟瞰视角下所述待检测目标的位置信息，所述主视偏差确定网络用于根据所述特征提取结果确定所述主视位置检测网络所确定的位置信息的偏差，所述鸟瞰偏差确定网络用于根据所述特征提取结果确定所述鸟瞰位置检测网络所确定的位置信息的偏差，所述姿态确定网络用于根据所述特征提取结果确定所述待检测目标的姿态信息。
  8. 根据权利要求1-7之一所述的方法，其中，所述待检测目标的数量为至少两个。
  9. 一种目标检测装置，其中，所述装置包括：
    图像获取单元，被配置成获取包含待检测目标的二维图像；
    检测单元，被配置成将所述图像输入至预先训练的目标检测网络，得到所述待检测目标的空间位置信息，其中，所述目标检测网络包括主视位置检测网络、鸟瞰位置检测网络和位置结合网络，所述主视位 置检测网络用于确定主视视角下所述待检测目标的位置信息，所述鸟瞰位置检测网络用于确定鸟瞰视角下所述待检测目标的位置信息，所述位置结合网络用于结合所述主视位置检测网络和鸟瞰位置检测网络分别确定的位置信息以得到所述待检测目标的空间位置信息。
  10. 根据权利要求9所述的装置，其中，所述目标检测网络还包括偏差确定网络和位置校正网络，所述偏差确定网络用于确定所述空间位置信息的偏差，所述位置校正网络用于根据所述空间位置信息的偏差对所述空间位置信息进行校正，得到校正后的空间位置信息。
  11. 根据权利要求10所述的装置，其中，所述偏差确定网络包括主视偏差确定网络、鸟瞰偏差确定网络和偏差结合网络，所述主视偏差确定网络用于确定所述主视位置检测网络所确定的位置信息的偏差，所述鸟瞰偏差确定网络用于确定所述鸟瞰位置检测网络所确定的位置信息的偏差，所述偏差结合网络用于结合所述主视偏差确定网络和鸟瞰偏差确定网络分别确定的偏差以得到所述空间位置信息的偏差。
  12. 根据权利要求11所述的装置，其中，所述装置还包括：
    三维重建单元，被配置成根据所述空间位置信息对所述待检测目标进行三维重建，得到三维重建结果。
  13. 根据权利要求12所述的装置，其中，所述装置还包括：
    确定单元，被配置成确定所述待检测目标的姿态信息；以及
    所述三维重建单元进一步被配置成根据空间位置信息和姿态信息，对待检测目标进行三维重建。
  14. 根据权利要求13所述的装置，其中，所述确定单元进一步被配置成：
    利用预先训练的姿态确定网络根据所述图像确定所述待检测目标 的姿态信息。
  15. 根据权利要求14所述的装置，其中，所述目标检测网络还包括特征提取网络，其中，所述特征提取网络用于提取所述图像的特征，得到特征提取结果；以及
    所述主视位置检测网络用于根据所述特征提取结果确定主视视角下所述待检测目标的位置信息，所述鸟瞰位置检测网络用于根据所述特征提取结果确定鸟瞰视角下所述待检测目标的位置信息，所述主视偏差确定网络用于根据所述特征提取结果确定所述主视位置检测网络所确定的位置信息的偏差，所述鸟瞰偏差确定网络用于根据所述特征提取结果确定所述鸟瞰位置检测网络所确定的位置信息的偏差，所述姿态确定网络用于根据所述特征提取结果确定所述待检测目标的姿态信息。
  16. 根据权利要求9-15之一所述的装置，其中，所述待检测目标的数量为至少两个。
  17. 一种电子设备，包括：
    一个或多个处理器；
    存储装置，其上存储有一个或多个程序；
    其中，当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如权利要求1-8中任一所述的方法。
  18. 一种计算机可读介质，其上存储有计算机程序，其中，该程序被处理器执行时实现如权利要求1-8中任一所述的方法。
PCT/CN2023/100388 2022-09-19 2023-06-15 目标检测方法和装置 WO2024060708A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211136508.5A CN115393423A (zh) 2022-09-19 2022-09-19 目标检测方法和装置
CN202211136508.5 2022-09-19

Publications (1)

Publication Number Publication Date
WO2024060708A1 true WO2024060708A1 (zh) 2024-03-28

Family

ID=84126162

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/100388 WO2024060708A1 (zh) 2022-09-19 2023-06-15 目标检测方法和装置

Country Status (2)

Country Link
CN (1) CN115393423A (zh)
WO (1) WO2024060708A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115393423A (zh) * 2022-09-19 2022-11-25 北京京东尚科信息技术有限公司 目标检测方法和装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111369617A (zh) * 2019-12-31 2020-07-03 浙江大学 一种基于卷积神经网络的单目视图的3d目标检测方法
US20200327690A1 (en) * 2019-04-09 2020-10-15 Sensetime Group Limited Three-dimensional object detection method and device, method and device for controlling smart driving, medium and apparatus
CN112668460A (zh) * 2020-12-25 2021-04-16 北京百度网讯科技有限公司 目标检测方法、电子设备、路侧设备和云控平台
CN113255520A (zh) * 2021-05-25 2021-08-13 华中科技大学 基于双目视觉和深度学习的车辆避障方法与电子设备
CN115393423A (zh) * 2022-09-19 2022-11-25 北京京东尚科信息技术有限公司 目标检测方法和装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200327690A1 (en) * 2019-04-09 2020-10-15 Sensetime Group Limited Three-dimensional object detection method and device, method and device for controlling smart driving, medium and apparatus
CN111369617A (zh) * 2019-12-31 2020-07-03 浙江大学 一种基于卷积神经网络的单目视图的3d目标检测方法
CN112668460A (zh) * 2020-12-25 2021-04-16 北京百度网讯科技有限公司 目标检测方法、电子设备、路侧设备和云控平台
CN113255520A (zh) * 2021-05-25 2021-08-13 华中科技大学 基于双目视觉和深度学习的车辆避障方法与电子设备
CN115393423A (zh) * 2022-09-19 2022-11-25 北京京东尚科信息技术有限公司 目标检测方法和装置

Also Published As

Publication number Publication date
CN115393423A (zh) 2022-11-25

Similar Documents

Publication Publication Date Title
WO2020006961A1 (zh) 用于提取图像的方法和装置
CN111079619B (zh) 用于检测图像中的目标对象的方法和装置
CN106846497B (zh) 应用于终端的呈现三维地图的方法和装置
US10970938B2 (en) Method and apparatus for generating 3D information
WO2020211573A1 (zh) 用于处理图像的方法和装置
EP3872764B1 (en) Method and apparatus for constructing map
CN112733820B (zh) 障碍物信息生成方法、装置、电子设备和计算机可读介质
CN109754464B (zh) 用于生成信息的方法和装置
CN109800730B (zh) 用于生成头像生成模型的方法和装置
US11453121B2 (en) Mirror loss neural networks
CN111784776B (zh) 视觉定位方法及装置、计算机可读介质和电子设备
CN110059623B (zh) 用于生成信息的方法和装置
WO2022033111A1 (zh) 图像信息提取方法、训练方法及装置、介质和电子设备
WO2020062494A1 (zh) 图像处理方法和装置
CN114399588B (zh) 三维车道线生成方法、装置、电子设备和计算机可读介质
WO2024060708A1 (zh) 目标检测方法和装置
CN111612842A (zh) 生成位姿估计模型的方法和装置
WO2020034981A1 (zh) 编码信息的生成方法和识别方法
CN111402122A (zh) 图像的贴图处理方法、装置、可读介质和电子设备
CN114998433A (zh) 位姿计算方法、装置、存储介质以及电子设备
CN112037305B (zh) 对图像中的树状组织进行重建的方法、设备及存储介质
CN113673446A (zh) 图像识别方法、装置、电子设备和计算机可读介质
CN111310595B (zh) 用于生成信息的方法和装置
CN109816791B (zh) 用于生成信息的方法和装置
CN115578432B (zh) 图像处理方法、装置、电子设备及存储介质

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: 23867004

Country of ref document: EP

Kind code of ref document: A1