WO2021082662A1 - 辅助用户拍摄车辆视频的方法及装置 - Google Patents

辅助用户拍摄车辆视频的方法及装置 Download PDF

Info

Publication number
WO2021082662A1
WO2021082662A1 PCT/CN2020/110735 CN2020110735W WO2021082662A1 WO 2021082662 A1 WO2021082662 A1 WO 2021082662A1 CN 2020110735 W CN2020110735 W CN 2020110735W WO 2021082662 A1 WO2021082662 A1 WO 2021082662A1
Authority
WO
WIPO (PCT)
Prior art keywords
shooting
current frame
image
vehicle
video
Prior art date
Application number
PCT/CN2020/110735
Other languages
English (en)
French (fr)
Inventor
郭昕
程远
王清
Original Assignee
支付宝(杭州)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 支付宝(杭州)信息技术有限公司 filed Critical 支付宝(杭州)信息技术有限公司
Publication of WO2021082662A1 publication Critical patent/WO2021082662A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N23/00—Cameras or camera modules comprising electronic image sensors; Control thereof
    • H04N23/60—Control of cameras or camera modules
    • H04N23/64—Computer-aided capture of images, e.g. transfer from script file into camera, check of taken image quality, advice or proposal for image composition or decision on when to take image

Definitions

  • One or more embodiments of this specification relate to the field of computer technology, and in particular to methods and devices for assisting users in shooting vehicle videos.
  • the method and device for model training using data from multiple data parties described in one or more embodiments of this specification can be used to solve one or more of the problems mentioned in the background art section.
  • a method for assisting a user in shooting a vehicle video includes: acquiring the current frame in the vehicle video shot by the user, and the current shooting state of a shooting terminal used to shoot the vehicle video Information; using a pre-trained image classification model to process the current frame, thereby obtaining image quality features of the current frame, and extracting shooting features from the shooting state information; at least inputting the image quality features into the pre-trained first detection Model to detect the validity of the current frame; in the case of detecting that the current frame is a valid frame, use a pre-trained component recognition model to identify vehicle components in the current frame to determine the vehicle component based on the recognition result The component feature of the current frame; using a pre-trained second detection model to process the shooting feature and the component feature to detect whether the current frame satisfies the preset shooting rule, so as to determine whether the current frame meets the preset shooting rule based on the detection result.
  • Frame of video shooting guide strategy includes: acquiring the current frame in the vehicle video shot by the user, and the current shooting state of
  • the current frame is an image frame extracted from the vehicle video at a predetermined time interval.
  • the photographing state information includes one or more of the following: acceleration magnitude, acceleration direction information, placement direction information, and position information of the photographing terminal.
  • the image quality feature includes at least one of the following: whether the image is clear, whether the image is a vehicle image, whether the light is sufficient, and whether the vehicle body is stained.
  • inputting at least the image quality feature into a pre-trained first detection model to detect the validity of the current frame includes: joining the image quality feature and the shooting feature and then inputting the pre-trained The first detection model is used to detect the validity of the current frame; wherein, the validity of the current frame also includes whether the image is shot upside down or shot upside down.
  • the method further includes: in a case where it is detected that the current frame is not a valid frame, providing an image shooting guide strategy for the current frame, the image shooting guide strategy includes one of the following : Shoot at the vehicle, shoot when the light is sufficient, shoot after cleaning the stains, and keep the shooting terminal in the vertical direction.
  • the shooting rule includes an image composition rule
  • the image composition rule indicates that a predetermined component falls into a predetermined area in the image
  • the video shooting guidance strategy includes adjusting the distance between the shooting terminal and the vehicle body.
  • the shooting rule includes a moving direction rule for detecting whether the moving direction of the photographing terminal is along a predetermined direction
  • the video shooting guide strategy includes moving in the opposite direction of the current moving direction, or returning to the origin for shooting.
  • the shooting rule includes a shooting angle rule for detecting whether the current frame crosses a predetermined shooting angle.
  • the second detection model is a long and short-term memory model.
  • an apparatus for assisting a user in shooting a vehicle video includes: an acquisition unit configured to acquire a current frame in the vehicle video shot by the user, and a shooting for shooting the vehicle video Current shooting state information of the terminal; a first feature extraction unit configured to process the current frame using a pre-trained image classification model, thereby acquiring image quality features of the current frame, and extracting shooting features from the shooting state information;
  • the validity detection unit is configured to input at least the image quality feature into a pre-trained first detection model to detect the validity of the current frame;
  • the second feature extraction unit is configured to detect that the current frame is valid
  • a pre-trained component recognition model is used to recognize vehicle components in the current frame to determine the component features of the current frame based on the recognition result;
  • the video shooting guidance unit is configured to use a pre-trained second detection model Processing uses the shooting feature and the component feature to detect whether the current frame satisfies a preset shooting rule, so as to determine a
  • a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method of the above-mentioned first aspect.
  • a computing device including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the above-mentioned first aspect is implemented. method.
  • the embodiments of this specification provide methods and devices for assisting users in shooting vehicle videos, making full use of the characteristics of the image itself and the shooting status information of the shooting terminal.
  • the validity of the image itself is detected to prevent the user from being captured in the vehicle video. Contains invalid frames, which affects the effect of vehicle inspection.
  • Figure 1 shows a schematic diagram of an implementation scenario of an embodiment of this specification
  • Fig. 2 shows a schematic diagram of a process of assisting a user in taking a video of a vehicle according to an embodiment
  • Figure 3 shows a specific example of a schematic diagram of judging the validity of the current frame
  • FIG. 4 shows a schematic diagram of a shooting direction, shooting angle, etc. in a shooting rule of a specific example
  • FIG. 5 shows a schematic diagram of a specific example of providing a video shooting guidance strategy
  • Fig. 6 shows a schematic block diagram of an apparatus for assisting a user in taking a video of a vehicle according to an embodiment
  • Fig. 1 shows a vehicle inspection scene, which can be any scene where the damage of the vehicle needs to be inspected. For example, when the vehicle is insured, it is determined whether the vehicle is damaged, or the vehicle is damaged during the auto insurance claims.
  • the user can collect live video of the vehicle through a mobile terminal that can collect video information (hereinafter referred to as a shooting terminal), such as a smart phone, a camera, and the like.
  • a shooting terminal such as a smart phone, a camera, and the like.
  • the processing platform can judge the validity of the currently collected image frame in real time, and provide auxiliary guidance for the user to shoot.
  • the processing platform may be integrated in a mobile terminal used by the user to collect video information, or may be located on a server that remotely provides services for vehicle inspection applications of the mobile terminal, which is not limited here.
  • the processing platform may obtain the current frame photographed by the user and the current state information of the photographing terminal. It can be understood that shooting features, such as shooting angle, movement acceleration, etc., can be extracted from the current state information.
  • shooting features such as shooting angle, movement acceleration, etc.
  • the quality of its quality can be judged through the pre-trained first detection model, for example, sharpness judgment, illuminance judgment, and so on.
  • the user may be prompted to adjust the shooting.
  • the component features of the current frame can be further extracted, combined with the shooting state features and component features, to detect whether the current frame meets the scene-related shooting rules, and based on the detection results, determine the video shooting for the user Guide strategy.
  • the predetermined conditions related to the scene are, for example, whether a predetermined component falls in a predetermined area in the image composition, whether the shooting angle change direction is a predetermined direction, and so on.
  • Fig. 2 shows a flowchart of a method for assisting a user in taking a video of a vehicle according to an embodiment.
  • the execution subject of the method can be any system, equipment, device, platform or server with computing and processing capabilities.
  • the method for assisting the user in shooting a vehicle video includes the following steps: Step 201, acquiring the current frame in the vehicle video shot by the user, and the current shooting state information of the shooting terminal used to shoot the vehicle video; Step 202, using The pre-trained image classification model processes the current frame, thereby obtaining the image quality characteristics of the current frame, and extracting the shooting features from the shooting status information; step 203, at least inputting the image quality features into the pre-trained first detection model to detect the current frame Validity; Step 204, in the case of detecting that the current frame is a valid frame, use a pre-trained component recognition model to identify vehicle components in the current frame from the current frame to determine the component features of the current frame based on the recognition result; step 205. Use a pre-trained second detection model to process the shooting state features and component features to detect whether the current frame meets a preset shooting rule, so as to determine a video shooting guidance strategy for the user based on the detection result.
  • step 201 the current frame in the vehicle video shot by the user and the current shooting state information of the shooting terminal used to shoot the vehicle video are acquired.
  • a certain shooting frame rate may be set, for example, 60 frames per second, that is, 60 frames per second. Since the process of executing the embodiment of this specification is synchronized with the shooting process, the real-time requirement of frame processing is relatively high. However, there may be a large number of overlapping areas in similar frames. Therefore, in this process, it is not required to process every frame in the captured video.
  • the acquired current frame may be the latest frame currently shot, or may be a frame that is determined to be basically synchronized with the currently acquired frame according to the processing performance of the device.
  • the shooting frame rate of the device is 60 frames per second, and according to the processing performance of the device, it can process 15 frames per second. Then, one frame can be extracted from every 4 frames at a predetermined frame interval (such as every 4 frames). The last frame), the process of this embodiment is executed. In this case, it may happen that when the 9th frame is captured, the current frame obtained is the 8th frame, which basically meets the real-time requirements. In the case that the processing performance of the device is good, and the real-time requirements can still be met by processing each frame, each frame can also be sequentially acquired as the current frame to execute the process of this embodiment.
  • the shooting state of the shooting terminal used to shoot the vehicle video is also important information in assisting the user in shooting the vehicle video. Therefore, in step 201, the current shooting state information of the shooting terminal can also be obtained.
  • the photographing terminal may be, for example, a terminal such as a smart phone or a camera.
  • the current shooting state information can be used to describe the current shooting state of the shooting terminal, for example, acceleration magnitude, acceleration direction, placement direction, position information, and so on.
  • the magnitude and direction of acceleration can be acquired through an acceleration sensor provided in the shooting terminal
  • the placement direction can be acquired through a gyroscope
  • the position information can be acquired through a positioning module (such as Beidouxing, GPS, etc.) implemented by software or hardware.
  • the placement direction may include vertical, horizontal (horizontal placement), tilt at a certain angle (upward shooting, overhead shooting), and so on.
  • the placement direction is related to the screen orientation of the captured image frame. For example, when the shooting terminal is a smart phone, due to the long body, the smart phone may be placed horizontally for shooting to achieve a better picture effect.
  • the position information may be absolute position information, for example, position information represented by latitude and longitude. If the absolute position error is large, the relative position of the photographing terminal can also be determined based on the data obtained through the gyroscope and the acceleration sensor.
  • the relative position may be a position relative to a reference position point.
  • the reference location point is, for example, a shooting starting point, that is, a location point where the shooting device is located when the first frame of the vehicle video is shot.
  • the frequency of changes in the shooting status of the shooting terminal is generally lower than the frequency of image collection. Therefore, the shooting status information may not be collected at every frame.
  • the shooting status information may be collected at a certain collection interval, for example, collected once every 0.5 seconds.
  • image quality features are features related to image quality.
  • the image quality feature may include, but is not limited to, at least one of the following: whether the image is clear, whether it is a photo of a vehicle, whether the vehicle body is stained, whether the light is sufficient, and so on.
  • the image classification model may be a multi-task classification model, for example, a model implemented by algorithms such as MobileNet V2 and ShuffleNet.
  • the image classification model can be trained through multiple images with annotated quality categories.
  • a picture can correspond to multiple tags.
  • a picture corresponds to labels such as "clear picture", "vehicle picture”, “no stains", "sufficient light”, etc.
  • the pictures used as training samples can be sequentially input to the selected model, and the model parameters can be adjusted based on the comparison between the output result of the model and the corresponding label.
  • the above sample labels can also be represented by numbers, for example, a clear picture corresponds to 1, and a blurred picture corresponds to 0.
  • the image classification model may include multiple output channels, and the current frame is input to the trained image classification model.
  • the image classification model may output the probability of the current frame in each category or the specific classification result through each channel.
  • the output results of the image classification model on the 4 channels are respectively 0.1, 0.8, 0.82, 0.9, which means that the blur probability of the picture is 0.1, and the probability of being a vehicle picture is 0.8.
  • the probability of stains is 0.82, and the probability of sufficient light is 0.9.
  • the output result of the image classification model on the 4 channels can also be 0, 1, 1, 1, indicating that the picture is clear, the picture of the vehicle, the body is stained, and the light is sufficient.
  • the shooting feature may be a feature related to the shooting state of the shooting terminal. For example, through the placement direction of the photographing terminal, it is possible to extract whether there are features such as upside-down shooting, upside-down shooting, etc. For example, when the angle formed by the shooting angle of the shooting terminal and the vertical upward direction is greater than a preset threshold (for example, 25 degrees), the shooting feature is determined to be overhead shooting, and so on. For another example, the direction of movement of the photographing terminal can be determined by its acceleration direction and speed.
  • a preset threshold for example 25 degrees
  • the image quality characteristics are input to the first pre-trained detection model to detect the validity of the current frame.
  • the current frame is valid, which means that the current frame is a frame that can be used to determine the state of the vehicle, for example, the image is clear, it is a vehicle picture, and so on.
  • the first detection model can be regarded as a classification model, for example, can be implemented as a model such as GBDT (Gradient Boosting Decision Tree).
  • the output result of the first detection model may correspond to a valid category and an invalid category.
  • the predetermined feature items are oriented toward the direction that can be used to judge the state of the vehicle. For example, whether the image is clear, the feature corresponds to the clear category, or the probability of clear picture Exceeds a predetermined value (such as 0.7) and so on.
  • the current frame can be judged as not a valid frame (may also be referred to as an invalid frame).
  • the detection result of the first detection model can correspond to multiple categories, which can be specific to various situations of the current frame, for example, when the current frame is not a valid frame, it is due to insufficient light. , The image is still not clear.
  • the first detection model can respectively correspond to various situations through the output results of multiple channels, and can also represent the current situation of the current frame through the output results of one channel.
  • each image quality feature can also have a priority. The priority is represented by the weight corresponding to the response feature in the model, for example, the larger the weight, the priority Higher.
  • the priority order is whether there is enough light, whether the image is clear, whether it is a photo of a vehicle, whether the body is stained, and so on.
  • This priority order also indicates the importance of the various factors of image quality. For example, only when sufficient light is detected and the image is clear can it be judged whether it is a picture of a vehicle.
  • the output result of the output channel of the detection model corresponds to the invalid situation or the valid frame situation with the highest priority.
  • the output result of the first detection model may also correspond to an image shooting guide strategy provided to the user according to the actual situation.
  • the image shooting guidance strategy is used to guide the user's image shooting behavior, which can include "please aim at the vehicle to shoot” (corresponding to non-vehicle images), "please shoot in a bright environment” (corresponding to insufficient light), “please scrub "Car body stains” (corresponding to body stains) and so on.
  • the output result of the first detection model may correspond to the image shooting guidance strategy of “please aim at the vehicle”.
  • the training sample of the first detection model can correspond to multiple pictures, and each picture contains the image quality characteristics determined by the image classification model, and pre-labeled classification labels, such as "please aim at the vehicle to shoot", “please in Shooting in a bright environment” and so on.
  • the image quality characteristics of each picture are respectively input to the selected classification model, and the model parameters are adjusted according to the comparison between the output result of the classification model and the classification label, so as to train the image classification model, which will not be repeated here.
  • the image quality feature and the shooting feature can also be stitched together and input into the first detection model, and the validity of the current frame can be detected in combination with the shooting feature.
  • some targets in the image may be deformed. For example, when shooting upside down, the target is pulled up, and the originally round target may become an ellipse. This is in part recognition and damage recognition. In the operation of the class, the accuracy of the result may be affected. Therefore, the shooting status of the shooting terminal can also be used as a factor in determining whether the current frame can determine the state of the vehicle.
  • the shooting feature of the shooting terminal includes overhead shooting or overhead shooting, it can be determined that the current frame is not a valid frame.
  • the output result of the first detection model can correspond to the invalid category, the overhead invalid category, the image shooting guidance strategy corresponding to the overhead invalid category (such as "please shoot vertically"), and so on.
  • the above image guidance strategy can be displayed to the user in the form of voice or text. In this way, if the current frame is not a valid frame, the user can be given an image shooting guide in time and collect a valid frame.
  • step 202 and step 203 are steps for detecting the validity of the current frame.
  • Fig. 3 is a schematic diagram of a specific example of judging the validity of a current frame.
  • the shooting state information of the shooting terminal can be collected every certain time interval (for example, 0.5 seconds), and updated to the current shooting state information.
  • the current shooting status information can be acquired.
  • the image quality characteristics of the current frame can be detected, represented by a feature vector, such as (1, 0, 0, 1), and the value on each element corresponds to the detection of the corresponding quality detection item result.
  • the shooting features such as acceleration features, gyroscope angle features, etc., are extracted from the current shooting state information, such as corresponding to (0.1, 30°).
  • the pre-trained component recognition model is used to identify the vehicle components in the current frame , To determine the feature of the current frame based on the recognition result.
  • the component recognition model can be used to recognize vehicle components in the image.
  • the part recognition model can be trained by using multiple pictures as training samples. These multiple pictures are respectively marked with part outlines and/or part names. During training, the original image can be input to the selected model, and the model parameters can be adjusted according to the annotation results, thereby training the component recognition model.
  • the component recognition model is realized by convolutional neural networks (CNN, Convolutional Neural Networks), for example.
  • the current frame when the current frame is a valid frame, the current frame may have the characteristics of clear image, sufficient light, vertical shooting, etc., input the current frame into the component recognition model, and the output result of the component recognition model can be marked with the contour of the component
  • the picture can also be a text description with component characteristics, such as the text describing the name and position relationship of each component appearing in the current frame.
  • the output result of the component recognition model can be used as the component feature of the currently photographed vehicle.
  • the shooting feature and the component feature are processed by using the pre-trained second detection model to detect whether the current frame meets the preset shooting rule, so as to determine the video shooting guidance strategy for the current frame based on the detection result.
  • the shooting rule may be a rule that determines that the relevant frame can fully display the state of the vehicle. According to actual needs, it can correspond to different shooting rules.
  • the shooting rules may include image composition rules.
  • the vehicle image composition rule may indicate that a predetermined part of the vehicle falls into a predetermined area of the image composition. For example, a predetermined area is indicated by a frame of a predetermined color, and when the current frame is shot, if the predetermined part falls within the frame, the image composition rule is satisfied.
  • a green frame can be given at the top and bottom of the image.
  • the chassis contour and the top contour can be determined according to the component characteristics.
  • the shooting rules may also include shooting angle rules.
  • shooting angle rules When images of various predetermined angles are collected, the state of the vehicle can be fully displayed from all directions. For example, if a circle of the car is divided into 12 angles, when the user collects images from all 12 angles, it can be considered that the user has collected a complete video of the vehicle.
  • the shooting angle can be determined according to the shooting characteristics. The shooting angle can be determined by the offset of the current shooting point relative to the reference point, the distance and the angle, for example.
  • the reference point may be the starting shooting position point.
  • the starting angle of view is the position directly in front of the vehicle
  • the frontal angle of view of the front license plate can be photographed, and the photographing characteristics of the current frame (such as current position point, image composition, etc.) can be taken at the beginning of the photographing.
  • the component features determine whether it is valid, and the shooting rules (such as whether the license plate appears in the middle of the screen in the left and right directions, etc.), if it is satisfied, it is an available frame.
  • a coordinate system can be established with the shooting position of the initial frame as the origin, and the longitudinal direction of the vehicle body and the lateral direction of the vehicle body as the coordinate axes.
  • the second detection model may include a calculation method for determining the relative angle.
  • the acceleration size and direction collected by the built-in acceleration sensor and gyroscope of the shooting device can be used to determine the angle of the shooting device relative to the starting point, or the current shooting angle of view relative to the starting point.
  • the angle of view is the angle through which the camera moves relative to the center axis (such as the center axis of the vehicle).
  • the gap between this angle and the preset angle is detected, and when the gap is within the preset gap range, the angle can be considered as the preset angle.
  • each preset angle can be detected in a circle around the body, it means that the shooting meets the shooting angle rule. Conversely, if a frame that meets a certain shooting angle rule is not detected within a gap of a preset angle, and the current frame has skipped the predetermined angle according to the moving direction of the shooting terminal, the current frame does not meet the shooting angle rule.
  • the integrity of the user's shooting angle can be controlled so as to obtain a comprehensive vehicle video.
  • the shooting rule may further include a movement direction rule. It can be understood that in the process of video shooting, moving and shooting in the same direction is more conducive to shooting and vehicle status determination.
  • the black wire frame around the car circle, and the arrow direction indicates the moving direction of the shooting terminal.
  • the second detection model may include an algorithm for determining the moving direction. The algorithm for determining the moving direction compares the photographing feature of the current frame with the photographing feature of the previous frame to determine the current moving direction of the photographing terminal.
  • the changing direction of the shooting position of the current frame relative to the shooting position of the previous frame is the moving direction of the shooting terminal.
  • the shooting rules are usually not only related to the current frame, but also related to the state of the previous frame. Therefore, in addition to the specific calculation methods described above, the second detection model can also be implemented by machine learning methods, for example, it can be a shallow layer. DNN model, LSTM model, etc. Taking LSTM as an example, it is suitable for dealing with timing problems, and can have memory and integration capabilities for long-term information. When the LSTM model is trained, multiple videos can be used as training samples. Each video frame can be extracted at a predetermined time interval, and each frame corresponds to an image feature, a shooting feature, and a label that meets the marked shooting rules.
  • the second detection model may further include multiple modules, each module detects one of the shooting rules, for example, one module is used to detect the shooting angle rule, one module is used to detect the image composition rule, and so on.
  • the output result of the second detection model can correspond to the satisfaction of the shooting rules, and can also correspond to the corresponding video shooting guidance strategy.
  • the video shooting guidance strategy can be, for example: when the moving direction is opposite, prompt the user to walk in the opposite direction; when the predetermined shooting angle is exceeded, prompt the user to return to the preset angle to shoot; when the image composition rules are not met, prompt the user to adjust the shooting terminal The distance to the body, and so on.
  • the video shooting guidance strategy may also be to return to the origin to shoot.
  • the current frame does not meet the pre-set image composition rules
  • the color of the frame of the predetermined color can be changed (for example, the green frame is converted to the red frame), and at the same time, the video can be prompted by the vibration of the mobile terminal, playing prompt music, etc.
  • a predetermined graphic such as an arrow, can also be used to indicate the correct moving direction.
  • FIG. 5 shows a schematic diagram of a specific example of detecting whether the current frame satisfies a preset shooting rule through the second detection model.
  • component identification is performed on the current frame to obtain component features.
  • the component feature and the shooting feature are stitched together to detect the moving direction of the shooting terminal and the image composition of the current frame using the two LSTM modules in the second detection model, respectively.
  • the moving direction is inconsistent with the prompt direction (if the direction of the arrow used in FIG.
  • a prompt can be given to the user to return to the original position on the front of the vehicle to take another shot.
  • the shooting composition is unreasonable, for example, the top of the vehicle does not fall into the set green frame, the user can be prompted to change the shooting mode by changing the frame color to red, shaking the camera, and voice prompts.
  • step 205 there is no need to provide a video shooting guide strategy, or the video shooting guide strategy is to continue shooting until the shooting is completed.
  • the shooting guidance strategy can always be displayed at the shooting terminal to guide the user in correct shooting.
  • the method for assisting the user in taking a video of a vehicle provided by the embodiment of this specification, on the one hand, for a single frame in the taken video, the validity as an image can be detected in real time. If a single frame is a valid frame, then the current frame is further used as a frame in the vehicle inspection video, whether it meets the vehicle inspection video shooting rules. In the case that the current frame is invalid or does not meet the vehicle inspection video shooting rules, the video shooting guide strategy can be provided to the user in time. In this way, ordinary users can correctly shoot effective vehicle inspection videos, which improves user experience and the efficiency of vehicle inspections during claims settlement. When the embodiment of this specification is used in the vehicle inspection process when the vehicle is insured, it can also reduce the claim risk of the insurance business party.
  • FIG. 6 shows a schematic block diagram of an apparatus for assisting a user in taking a video of a vehicle according to an embodiment.
  • the device can be installed in equipment, platforms, terminals, and servers with certain computing capabilities, such as a smart phone as shown in FIG. 1. As shown in FIG. 1,
  • the device 600 for assisting a user in shooting a vehicle video includes: an acquiring unit 61 configured to acquire the current frame in the vehicle video shot by the user and the current shooting state information of the shooting terminal used to shoot the vehicle video; first The feature extraction unit 62 is configured to use a pre-trained image classification model to process the current frame, thereby acquiring image quality features of the current frame, and extract the shooting features from the shooting status information; the validity detection unit 63 is configured to input at least the image quality features A pre-trained first detection model to detect the validity of the current frame; the second feature extraction unit 64 is configured to use the pre-trained component recognition model to identify the vehicle in the current frame when the current frame is detected as a valid frame Component to determine the component feature of the current frame based on the recognition result; the video shooting guide unit 65 is configured to use a pre-trained second detection model to process the shooting feature and component feature to detect whether the current frame meets the preset shooting rule, Therefore, a video shooting guidance strategy for the current frame is determined based on the detection result.
  • the current frame is an image frame extracted from the vehicle video at a predetermined time interval.
  • the shooting status information includes one or more of the following: acceleration magnitude and acceleration direction information acquired by an acceleration sensor provided in the shooting terminal, placement direction information acquired by a gyroscope, and information acquired by a positioning module location information.
  • the image quality feature includes at least one of the following: whether the image is clear, whether the image is a vehicle image, whether the light is sufficient, and whether the vehicle body is stained.
  • the validity detection unit 63 is further configured to: input the pre-trained first detection model after stitching the image quality features and the shooting features to detect the validity of the current frame; wherein, the validity of the current frame also includes: Whether to take the image upside down or upside down.
  • the apparatus 600 further includes: an image shooting guide unit (not shown) configured to provide an image shooting guide strategy for the current frame in the case of detecting that the current frame is not a valid frame.
  • the image shooting guide strategy includes the following One of the following: shooting at the vehicle, shooting when the light is sufficient, shooting after cleaning the stains, and keeping the shooting terminal in the vertical direction.
  • the shooting rule includes an image composition rule
  • the image composition rule indicates that a predetermined component falls into a predetermined area in the image
  • the video shooting guidance strategy includes adjusting the distance between the shooting terminal and the vehicle body.
  • the shooting rule includes a moving direction rule for detecting whether the moving direction of the photographing terminal is in a predetermined direction
  • the video shooting guide strategy includes moving in the opposite direction of the current moving direction, or returning to the origin to shoot.
  • the shooting rule includes a shooting angle rule for detecting whether the current frame crosses a predetermined shooting angle.
  • the above device 600 for assisting the user in shooting vehicle video shown in FIG. 6 corresponds to the method embodiment shown in FIG. 2, and the corresponding description in the method embodiment corresponding to FIG. 2 is also applicable to the method shown in FIG. 6
  • the device shown to assist the user in taking the video of the vehicle will not be repeated here.
  • a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the correspondingly described method.
  • a computing device including a memory and a processor, the memory stores executable code, and the processor implements the correspondingly described method when the executable code is executed.

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Studio Devices (AREA)
  • Image Analysis (AREA)

Abstract

辅助用户拍摄车辆视频的方法和装置，一方面，对于所拍摄的视频中的单个帧，可以实时检测其作为图像的有效性。如果单个帧是有效帧，则进一步判断对于该当前帧作为验车视频中的帧，是否符合验车视频拍摄规则。在当前帧无效，或者不符合验车视频拍摄规则的情况下，可以及时向用户提供拍摄引导策略。如此，可以使得普通用户能够正确拍摄有效的验车视频，提高用户体验，以及验车效率。

Description

辅助用户拍摄车辆视频的方法及装置 技术领域
本说明书一个或多个实施例涉及计算机技术领域，尤其涉及辅助用户拍摄车辆视频的方法及装置。
背景技术
在传统车险理赔场景中，往往通过保险业务方的专业查勘人员进行验车。由于需要人工到事故现场进行查勘定损，保险业务方需要投入大量的人力成本，和专业知识的培训成本。从普通用户的体验来说，理赔流程由于等待人工查勘员现场查验等，用户的等待时间较长，体验较差。而在车险投保场景中，通常没有验车环节。这样就可能出现带伤投保等情况，保险公司存在较大的理赔风险。
针对以上需求背景，开始设想将人工智能和机器学习应用到车辆损伤检测的场景中，希望能够利用人工智能领域计算机视觉图像识别技术，根据普通用户拍摄的现场图像(图片或视频)，在车险投保或理赔等场景中自动识别其中反映的车辆状况。如此，可以大大减少人工成本，同时提升用户体验。然而，普通用户自主拍摄现场图像过程中，如果拍摄过程或者所拍摄的图像存在不规范问题，那么会大大增加保险业务方的承保风险。
发明内容
本说明书一个或多个实施例描述的利用多个数据方的数据进行模型训练的方法及装置，可以用于解决背景技术部分提到的一个或多个问题。
根据第一方面，提供了一种辅助用户拍摄车辆视频的方法，其中，所述方法包括：获取用户拍摄的车辆视频中的当前帧，以及用于拍摄所述车辆视频的拍摄终端的当前拍摄状态信息；利用预先训练的图像分类模型处理所述当前帧，从而获取所述当前帧的图像质量特征，从所述拍摄状态信息提取拍摄特征；至少将所述图像质量特征输入预先训练的第一检测模型，以检测所述当前帧的有效性；在检测到所述当前帧是有效帧的情况下，利用预先训练的部件识别模型识别所述当前帧中的车辆部件，以基于识别结果确定所述当前帧的部件特征；利用预先训练的第二检测模型处理通过所述拍摄特征和所述部件特征，以检测所述当前帧是否满足预先设定的拍摄规则，从而基于检测结果确定针对 所述当前帧的视频拍摄引导策略。
在一个实施例中，所述当前帧是从所述车辆视频中按照预定时间间隔抽取的图像帧。
在一个实施例中，所述拍摄状态信息包括以下中的一项或多项：所述拍摄终端的加速度大小、加速度方向信息、放置方向信息、位置信息。
在一个实施例中，所述图像质量特征包括以下中的至少一项：图像是否清晰、图像是否为车辆图像、光线是否充足、车身是否有污渍。
在一个实施例中，至少将所述图像质量特征输入预先训练的第一检测模型，以检测所述当前帧的有效性包括：将所述图像质量特征和所述拍摄特征拼接后输入预先训练的第一检测模型，以检测所述当前帧的有效性；其中，所述当前帧的有效性还包括，是否仰拍图像或俯拍图像。
在一个实施例中，所述方法还包括：在检测到所述当前帧不是有效帧的情况下，提供针对所述当前帧的图像拍摄引导策略，所述图像拍摄引导策略包括以下中的一项：对准车辆拍摄、在光线充足时拍摄、清洗污渍后拍摄、保持拍摄终端沿竖直方向。
在一个实施例中，所述拍摄规则包括图像构图规则，所述图像构图规则指示预定部件落入图像中的预定区域，所述视频拍摄引导策略包括调整所述拍摄终端与车身的距离。
在一个实施例中，所述拍摄规则包括移动方向规则，用于检测拍摄终端的移动方向是否沿预定方向，所述视频拍摄引导策略包括向当前移动方向的相反方向移动，或者返回原点拍摄。
在一个实施例中，所述拍摄规则包括拍摄角度规则，用于检测所述当前帧是否跨越预定拍摄角度。
在一个实施例中，所述第二检测模型是长短期记忆模型。
根据第二方面，提供了一种辅助用户拍摄车辆视频的装置，其中，所述装置包括：获取单元，配置为获取用户拍摄的车辆视频中的当前帧，以及用于拍摄所述车辆视频的拍摄终端的当前拍摄状态信息；第一特征提取单元，配置为利用预先训练的图像分类模型处理所述当前帧，从而获取所述当前帧的图像质量特征，并从所述拍摄状态信息提取拍摄特征；有效性检测单元，配置为至少将所述图像质量特征输入预先训练的第一检测 模型，以检测所述当前帧的有效性；第二特征提取单元，配置为在检测到所述当前帧是有效帧的情况下，利用预先训练的部件识别模型识别所述当前帧中的车辆部件，以基于识别结果确定所述当前帧的部件特征；视频拍摄引导单元，配置为利用预先训练的第二检测模型处理通过所述拍摄特征和所述部件特征，以检测所述当前帧是否满足预先设定的拍摄规则，从而基于检测结果确定针对所述当前帧的视频拍摄引导策略。
根据第三方面，提供了一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算机中执行时，令计算机执行上述第一方面的方法。
根据第四方面，提供了一种计算设备，包括存储器和处理器，其特征在于，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现上述第一方面的方法。
本说明书实施例提供了辅助用户拍摄车辆视频的方法和装置，充分利用图像本身的特征，和拍摄终端的拍摄状态信息，一方面，对图像本身的有效性进行检测，以免用户拍摄的车辆视频中包含无效帧，影响验车效果，另一方面，检测各个帧是否满足拍摄规则，避免拍摄的车辆视频无法全面展示车辆状态。如此，通过两个方面为用户提供拍摄引导，可以提高用户拍摄的车辆视频的有效性。
附图说明
为了更清楚地说明本申请实施例的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1示出本说明书实施例的一个实施场景示意图；
图2示出根据一个实施例的辅助用户拍摄车辆视频的流程示意图；
图3示出一个具体例子的判断当前帧有效性的示意图；
图4示出一个具体例子的拍摄规则中拍摄方向、拍摄角度等的示意图；
图5示出一个具体例子的提供视频拍摄引导策略的示意图；
图6示出根据一个实施例的辅助用户拍摄车辆视频的装置的示意性框图
具体实施方式
下面结合附图，对本说明书提供的方案进行描述。
首先，结合图1示出一个具体实施场景进行说明。如图1所示，在该实施场景中，图1示出的是在验车场景，该场景可以是需要对车辆损伤情况进行查验的任意场景。例如，车辆投保时确定车辆有无损伤，或者车险理赔时确定车辆损伤情况等等场景。
该实施场景中，用户可以通过可采集视频信息的移动终端(以下称为拍摄终端)，例如智能手机、照相机等，采集车辆的现场视频。在用户采集车辆的现场视频过程中，处理平台可以实时判断当前采集的图像帧的有效性，对用户拍摄提供辅助引导。其中，处理平台可以是集成在用户用于采集视频信息的移动终端，也可以设在远程为移动终端的验车类应用提供服务的服务端，在此不作限定。
具体地，处理平台可以获取用户拍摄的当前帧，以及拍摄终端的当前状态信息。可以理解，从当前状态信息中可以提取拍摄特征，例如拍摄角度、移动加速度等。一方面，对于当前帧，可以通过预先训练的第一检测模型对其质量进行有效性判断，例如，清晰度判断、光照度判断等等。在当前帧的质量存在问题判断为无效帧时，可以提示用户调整拍摄。另一方面，在当前帧判断为有效帧时，可以进一步提取当前帧的部件特征，结合拍摄状态特征和部件特征，检测当前帧是否满足场景相关的拍摄规则，并基于检测结果对用户确定视频拍摄引导策略。其中，场景相关的预定条件例如是在图像构图中预定部件是否落在预定区域，拍摄角度改变方向是否是预定方向，等等。
下面详细描述辅助用户拍摄车辆视频的具体过程。
图2示出根据一个实施例的辅助用户拍摄车辆视频的方法流程图。该方法的执行主体可以是任何具有计算、处理能力的系统、设备、装置、平台或服务器。例如图1示出的处理平台。
如图2所示，辅助用户拍摄车辆视频的方法包括以下步骤：步骤201，获取用户拍摄的车辆视频中的当前帧，以及用于拍摄车辆视频的拍摄终端的当前拍摄状态信息；步骤202，利用预先训练的图像分类模型处理当前帧，从而获取当前帧的图像质量特征，并从拍摄状态信息中提取拍摄特征；步骤203，至少将图像质量特征输入预先训练的第一检测模型，以检测当前帧的有效性；步骤204，在检测到当前帧是有效帧的情况下，利用预先训练的部件识别模型从当前帧中识别当前帧中的车辆部件，以基于识别结果确定当前帧的部件特征；步骤205，利用预先训练的第二检测模型处理拍摄状态特征和部 件特征，以检测当前帧是否满足预先设定的拍摄规则，从而基于检测结果确定针对用户的视频拍摄引导策略。
首先，在步骤201中，获取用户拍摄的车辆视频中的当前帧，以及用于拍摄车辆视频的拍摄终端的当前拍摄状态信息。
其中，用户拍摄车辆视频时，根据设备的性能，可以有一定的拍摄帧率，例如60帧/秒，即每秒钟拍摄60帧。由于执行本说明书实施例的流程，是与拍摄过程同步的，因此，帧处理的实时性要求较高。而相近帧可能存在大量重复区域，因此，在本流程中，不要求对所拍摄的视频中的每一帧都进行处理。本步骤201中，所获取的当前帧，可以是当前拍摄的最新一帧，也可以是按照设备的处理性能，确定的可以与当前采集的帧基本同步的帧。例如，设备拍摄帧率为60帧/秒，而按照设备的处理性能，每秒可以处理15帧，那么，可以按照预定的帧间隔，从每4帧中抽取一帧(如每4帧中的最后一帧)，执行本实施例的流程。这种情况下，可能会出现拍摄到第9帧时，获取的当前帧是第8帧的情况，基本满足实时性要求。在设备的处理性能较好，对每一帧都进行处理仍可以满足实时性要求的情况下，还可以按顺序获取每一帧作为当前帧执行本实施例的流程。
另一方面，用于拍摄车辆视频的拍摄终端的拍摄状态也是辅助用户拍摄车辆视频中的重要信息，因此，在步骤201中还可以获取拍摄终端的当前拍摄状态信息。拍摄终端例如可以是智能手机、照相机之类的终端。当前拍摄状态信息可以用于描述拍摄终端当前拍摄所处的状态，例如，加速度大小、加速度方向、放置方向、位置信息等等。其中，加速度大小和加速度方向可以通过设于拍摄终端的加速度传感器获取，放置方向可以通过陀螺仪获取，位置信息可以通过软件或硬件实现的定位模块(如北斗星、GPS等)获取。
值得说明的是，放置方向可以包括竖直、水平(横向放置)、倾斜某个角度(仰拍、俯拍)，等等。放置方向与拍摄的图像帧的画面方向相关，例如，当拍摄终端是智能手机时，由于车身较长，可能将智能手机横向放置拍摄以达到更好的画面效果。
位置信息可以是绝对位置信息，例如通过经纬度表示的位置信息。如果采用绝对位置误差较大，还可以基于通过陀螺仪、加速度传感器获取数据确定拍摄终端的相对位置。相对位置可以是相对于参考位置点的位置。参考位置点例如是拍摄起点，即拍摄车辆视频的第一帧时拍摄设备所在的位置点。
可以理解，拍摄终端的拍摄状态变化频率通常低于图像采集频率，因此，拍摄状 态信息不一定在采集每一帧时都采集。可选地，拍摄状态信息可以是按照一定的采集间隔进行采集的，例如，每0.5秒采集一次。
接着，通过步骤202，利用预先训练的图像分类模型处理当前帧，从而获取当前帧的图像质量特征，并从拍摄状态信息中提取拍摄特征。其中，图像质量特征是与图像质量相关的特征。这里，图像质量特征例如可以包括但不限于以下至少一项：图像是否清晰、是否为车辆照片、车身是否有污渍、光线是否充足，等等。
图像分类模型可以是一个多任务分类模型，例如通过诸如MobileNet V2，ShuffleNet之类的算法实现的模型。图像分类模型可以通过被标注质量类别的多张图片进行训练。一张图片可以对应多个标签。例如，一张图片对应着“清晰图片”、“车辆图片”、“无污渍”、“光线充足”等等标签。对图像分类模型进行训练时，可以将作为训练样本的图片依次输入选定的模型，并基于模型的输出结果和相应标签的对比，调整模型参数。上述的样本标签也可以用数字表示，例如清晰图片对应1，模糊图片对应0。
图像分类模型可以包括多个输出通道，将当前帧输入训练好的图像分类模型，图像分类模型可以通过各个通道输出当前帧在各个分类上的概率，或者具体分类结果。作为示例，假设当前帧输入图像分类模型后，图像分类模型在4个通道上的输出结果分别是0.1，0.8，0.82，0.9，表示图片模糊概率为0.1，是车辆图片的概率为0.8，车身有污渍的概率为0.82，光线充足概率为0.9。如果预先确定有截断概率，图像分类模型在4个通道上的输出结果还可以是0，1，1，1，表示图片清晰、是车辆图片、车身有污渍、光线充足。
这样，可以获得出与图像的质量相关的特征。其中，在使用MobileNet V2，ShuffleNet等算法时，可以对传统的二维卷积操作进行优化，很好地减少模型参数，加快运算效率，便于在移动端部署，在此不再赘述。
另一方面，还可以从拍摄状态信息中提取出拍摄特征。顾名思义，拍摄特征可以是与拍摄终端的拍摄状态相关的特征。例如，通过拍摄终端的放置方向，可以提取是否存在仰拍、俯拍等状态的特征。如拍摄终端的拍摄角度和竖直向上方向所成角度大于预设阈值(如25度)时，确定拍摄特征为仰拍，等等。再例如，通过拍摄终端的加速度方向和速度可以确定其移动方向。
进一步地，在步骤203中，至少将图像质量特征输入预先训练的第一检测模型， 以检测当前帧的有效性。可以理解，当前帧有效，表示当前帧是可以用于判断车辆状态的帧，例如图像清晰，是车辆图片，等等。第一检测模型可以看作一个分类模型，例如可以实现为诸如GBDT(梯度提升决策树)之类的模型。
在一些可选的实现方式中，第一检测模型的输出结果可以对应有效类别和无效类别。在第一检测模型的输出结果对应到有效类别时，至少预定的特征项都是向着可以用于判断车辆状态的方向的，例如图像是否清晰的特征对应的是清晰类别，或者是清晰图片的概率超过预定值(如0.7)等等。通常，光线不充足、图像不清晰、不是车辆照片、车身有污渍等等中的至少一种情况下，当前帧可以被判断为不是有效帧(也可以称为无效帧)。
在另一些可选的实现方式中，第一检测模型的检测结果可以对应到多个类别，这多个类别可以具体到当前帧的各种情形，例如当前帧不是有效帧时，是光线不充足，还是图像不清晰。第一检测模型可以通过多个通道的输出结果分别对应各种情形，也可以通过一个通道的输出结果表示当前帧的当前情形。当第一检测模型通过一个通道的输出结果表示当前帧的当前情形时，各个图像质量特征还可以有优先级，优先级例如通过模型中响应特征对应的权重来表示，权重越大，表示优先级越高。例如，优先级顺序为光线是否充足、图像是否清晰、是否为车辆照片、车身是否有污渍，等等。这个优先级顺序也表示出图像质量的各个因素之间的重要度。例如，只有检测到光线充足，且图像清晰时，才可以判断是不是车辆图片。检测模型的输出通道的输出结果对应优先级别最高的无效情形，或者有效帧情形。
在一个实施例中，如果当前帧不是有效帧，第一检测模型的输出结果还可以对应着根据实际情形为用户提供的图像拍摄引导策略。其中，图像拍摄引导策略用于引导用户的图像拍摄行为，其可以包括“请对准车辆拍摄”(对应非车辆图像)、“请在明亮环境下拍摄”(对应光线不充足)、“请擦洗车身污渍”(对应车身有污渍)等等。例如，检测到是否车辆图像特征对应着非车辆，则第一检测模型的输出结果可以对应着“请对准车辆拍摄”的图像拍摄引导策略。
这时，第一检测模型的训练样本可以对应着多张图片，每张图片包含经过图像分类模型确定的图像质量特征，以及预先标注的分类标签，如“请对准车辆拍摄”、“请在明亮环境下拍摄”等等。将各个图片的图像质量特征分别输入选定的分类模型，根据分类模型的输出结果与分类标签的对比调整模型参数，从而训练图像分类模型，在此不再赘述。
根据一个可能的设计，还可以将图像质量特征和拍摄特征经过拼接后一起输入第一检测模型，结合拍摄特征检测当前帧的有效性。可以理解，对于特殊角度拍摄的图像帧，图像中的一些目标可能会产生变形，例如仰拍时，目标被拉高，本来是圆形的目标可能变成椭圆，这在部件识别、损伤识别之类的操作中，可能影响结果的准确度。因此，拍摄终端的拍摄状态也可以作为当前帧能否判断车辆状态的因素。例如在一个实施例中，如果拍摄终端的拍摄特征包括仰拍或俯拍，则可以判断当前帧不是有效帧。此时，根据前述的不同实施例，第一检测模型的输出结果可以对应到无效类别、俯拍无效类别、俯拍无效类别对应的图像拍摄引导策略(如“请竖直拍摄”)等等。
上面的图像引导策略可以以语音或文字等形式展示给用户，如此，如果当前帧不是有效帧，可以及时给出用户图像拍摄引导，采集有效帧。
有以上描述可知，步骤202和步骤203是用于对当前帧的有效性进行检测的步骤。为了更清楚地描述以上过程，请参考图3所示。图3是判断一个当前帧的有效性的具体例子示意图。在图3中，当拍摄终端开始拍摄车辆视频后，可以每间隔一定时间(如0.5秒)采集拍摄终端的拍摄状态信息，并更新为当前拍摄状态信息。在获取当前帧的同时，可以获取当前的拍摄状态信息。然后，一方面，通过多任务分类模型，可以检测当前帧的图像质量特征，用特征向量表示，例如(1，0，0，1)，每个元素上的值对应着相应质量检测项的检测结果。另一方面，通过当前拍摄状态信息提取拍摄特征，例如加速度特征、陀螺仪角度特征等等，如对应着(0.1，30°)。接着，把图像质量特征和拍摄特征进行拼接，得到(1，0，0，1，0.1，30°)，输入预先训练的GBDT模型(第一检测模型)，根据GBDT模型的输出结果确定当前帧是有效帧，或者对应的无效情形下的提示信息，如夜间(光线不足)情形下的“请在明亮环境下拍摄”、仰拍/俯拍情形下的“请竖直拍摄”等等。
在第一检测模型的输出结果对应到当前帧是有效帧的情况下，即当前帧可以用于判断车辆状态的帧，则通过步骤204，利用预先训练的部件识别模型识别当前帧中的车辆部件，以基于识别结果确定当前帧的部件特征。其中，部件识别模型可以用于识别图像中的车辆部件。
部件识别模型可以通过多张图片作为训练样本进行训练。这多张图片分别标注有部件轮廓和/或部件名称。进行训练时，可以将原始图片输入选定的模型，根据标注结果调整模型参数，从而训练部件识别模型。部件识别模型例如通过卷积神经网络(CNN，Convolutional Neural Networks)等实现。
根据步骤203的描述，当前帧为有效帧时，当前帧可能具有图像清晰、光线充足、竖直拍摄等特征，将当前帧输入部件识别模型，部件识别模型的输出结果可以是标注有部件轮廓的图片，也可以是具有部件特征的文字描述，如文字描述出当前帧中出现的各个部件的名称及位置关系等。部件识别模型的输出结果，可以作为当前拍摄车辆的部件特征。
然后，通过步骤205，利用预先训练的第二检测模型处理拍摄特征和部件特征，以检测当前帧是否满足预先设定的拍摄规则，从而基于检测结果确定针对当前帧的视频拍摄引导策略。其中，拍摄规则可以是确定相关帧可以完整展示车辆状态的规则。根据实际需求，可以对应不同的拍摄规则。
在一个实施例中，拍摄规则可以包括图像构图规则。车辆图像构图规则可以指示车辆的预定部位落入图像构图的预定区域。例如，通过预定颜色的框指示出预定区域，拍摄当前帧时，如果预定部位落入该框，则满足图像构图规则。作为示例，可以在图像顶部和底部分别给定一个绿色的框，拍摄过程中各个帧的车辆轮廓中，底盘轮廓需要落入底部的框，同时，顶端轮廓需要落入顶部的框，否则，当前帧不满足图像构图规则。其中，底盘轮廓和顶部轮廓可以根据部件特征确定。通过图像构图规则，可以使得用户与车身距离均衡，避免拍摄过程中忽远忽近，视频帧中的车辆图像忽大忽小等情形。
在另一个实施例中，拍摄规则还可以包括拍摄角度规则。当各个预定角度的图像被采集时，可以从各个方位完整展示车辆状态。例如，将环车一周分为12个角度，当用户在这12个角度都进行了图像采集，就可以认为用户采集了完整的车辆视频。其中，拍摄角度可以根据拍摄特征确定。拍摄角度例如可以通过当前拍摄点相对于参考点，距离和角度的偏移量来确定。可选地，参考点可以是起始拍摄位置点。
为了便于描述，请参考图4示出的示例。在图4中，假设起始视角是正对车辆前方的位置，可以拍摄到前方车牌的正面的视角，则在起始拍摄时，可以根据当前帧的拍摄特征(如当前位置点、图像构图等)以及部件特征确定是否有效，以及满足拍摄规则(如车牌是否出现在左右方向的画面中间位置等)，如果满足，就是可用的帧。起始帧确定后，可以以起始帧的拍摄位置为原点，车身纵向方向和车身横向方向为坐标轴建立坐标系。
在一个可选的实现方式中，第二检测模型可以包含确定相对角度的计算方法。当拍摄终端移动时，通过拍摄设备内置的加速度传感器、陀螺仪等采集的加速度大小和方向，利用相对角度的计算方法，可以确定拍摄设备相对于起始点的角度，或者当前拍摄 视角相对于起始视角沿拍摄设备移动相对的中心轴(如车辆中心轴线)转过的角度。检测这个角度和预设的角度之间的差距，当该差距在预设的差距范围内时，可以认为这个角度是预设角度。当环绕车身一周可以检测到各个预设的角度，就代表着拍摄满足拍摄角度规则。反之，如果在一个预设角度的差距范围内没有检测到满足某个拍摄角度规则的帧，而当前帧已经按照拍摄终端移动方向跳过了这个预定角度，则当前帧不满足拍摄角度规则。
通过拍摄角度规则，可以控制用户的拍摄角度的完整性，以便获取全面的车辆视频。
在又一个实施例中，拍摄规则还可以包括移动方向规则。可以理解，在视频拍摄过程中，按照同一个方向移动拍摄是较有利于拍摄和车辆状态认定的，如图4所示的环车一周的黑色线框，箭头方向表示拍摄终端的移动方向。可选地，第二检测模型中可以包括移动方向的确定算法。移动方向的确定算法将当前帧的拍摄特征与前一帧的拍摄特征的对比，可以确定拍摄终端的当前移动方向。作为示例，当前帧的拍摄位置相对于前一帧的拍摄位置的变化方向，就是拍摄终端的移动方向。
在更多实施例中，还可以有更多拍摄规则，在此不再赘述。
通过以上描述可知，拍摄规则通常不仅和当前帧相关，还和前一帧的状态相关，因此，除了上述的具体计算方法，第二检测模型还可以通过机器学习方法来实现，例如可以是浅层DNN模型、LSTM模型等等。以LSTM为例，适用于处理时序问题，可以对长时间的信息具有记忆能力，以及整合能力。LSTM模型训练时，可以将多个视频作为训练样本。每个视频中可以按照预定时间间隔抽取各个帧，每个帧对应有图像特征和拍摄特征，以及经过标注的拍摄规则的满足情况的标签。这些标签例如是：满足拍摄规则、移动方向相反、越过预定拍摄角度等等。针对各个训练样本，按照时间顺序依次将各个帧的图像特征和拍摄特征输入LSTM模型，并根据当前帧的标注结果调整模型参数，从而训练LSTM模型。可选地，第二检测模型还可以包括多个模块，每个模块检测其中一个拍摄规则，例如，一个模块用于检测拍摄角度规则，一个模块用于检测图像构图规则，等等。
第二检测模型的输出结果可以对应到对拍摄规则的满足情况，也可以对应到相应的视频拍摄引导策略。视频拍摄引导策略例如可以是：在移动方向相反时，提示用户向相反方向行走；越过预定拍摄角度时，提示用户回到预设角度范围内拍摄；不符合图像构图规则时，提示用户调整拍摄终端与车身的距离，等等。可选地，当移动方向不正确 时，视频拍摄引导策略还可以是返回原点拍摄。
向用户提供视频拍摄策略时，可以通过语音、文字、图像等等各种方式。例如，如果当前帧不满足预先设定的图像构图规则，预定颜色的框可以转变颜色(如绿色框转变成红色框)，同时，还可以通过移动终端的震动、播放提示音乐等方式提示视频拍摄存在的问题。在当前帧不满足预先设定的移动方向规则等时，还可以通过预定图形，如箭头，指示正确移动方向等。
为了更直观描述步骤204和步骤205的实施方案，请参考图5所示。图5示出了通过第二检测模型检测当前帧是否满足预先设定的拍摄规则的具体例子的示意图。如图5所示的具体例子中，在检测到当前帧是有效帧的情况下，对当前帧进行部件识别，得到部件特征。然后，将部件特征和拍摄特征拼接分别使用第二检测模型中的两个LSTM模块对拍摄终端的移动方向和当前帧的图像构图进行检测。当移动方向与提示方向不一致(如遇图4使出的箭头方向相反)，可以给出用户回到车辆正面的原点位置重新拍摄的提示。另一方面，当拍摄构图不合理时，例如车辆顶部没落入设定的绿色框，则可以通过将框颜色变成红色、拍摄终端震动、语音提示等方式提示用户改变拍摄方式。
这样，如果当前帧无效，可以及时在步骤203中确定当前帧无效后为用户提供图像拍摄引导，如果当前帧有效，则在步骤205中为用户提供视频拍摄引导，从而避免用户的不规范拍摄操作导致的后续问题。可以理解的是，如果用户操作规范，步骤205中，无需提供视频拍摄引导策略，或者说视频拍摄引导策略为继续拍摄，直至拍摄完成。可选地，不管用户拍摄是否规范，都可以一直在拍摄终端展示拍摄引导策略，以引导用户的正确拍摄。
回顾以上过程，本说明书实施例所提供的辅助用户拍摄车辆视频的方法，一方面，对于所拍摄的视频中的单个帧，可以实时检测其作为图像的有效性。如果单个帧是有效帧，则进一步对于该当前帧作为验车视频中的帧，是否符合验车视频拍摄规则。在当前帧无效，或者不符合验车视频拍摄规则的情况下，可以及时向用户提供视频拍摄引导策略。如此，可以使得普通用户能够正确拍摄有效的验车视频，提高用户体验，以及理赔时的验车效率。在本说明书实施例用于车辆投保时的验车流程时，还可以降低保险业务方的理赔风险。
根据另一方面的实施例，还提供一种辅助用户拍摄车辆视频的装置。图6示出根据一个实施例的辅助用户拍摄车辆视频的装置的示意性框图。该装置可以设置于有一定计算能力的设备、平台、终端、服务器，如图1示出的智能手机。如图6所示，辅助用 户拍摄车辆视频的装置600包括：获取单元61，配置为获取用户拍摄的车辆视频中的当前帧，以及用于拍摄车辆视频的拍摄终端的当前拍摄状态信息；第一特征提取单元62，配置为利用预先训练的图像分类模型处理当前帧，从而获取当前帧的图像质量特征，并从拍摄状态信息提取拍摄特征；有效性检测单元63，配置为至少将图像质量特征输入预先训练的第一检测模型，以检测当前帧的有效性；第二特征提取单元64，配置为在检测到当前帧是有效帧的情况下，利用预先训练的部件识别模型识别当前帧中的车辆部件，以基于识别结果确定当前帧的部件特征；视频拍摄引导单元65，配置为利用预先训练的第二检测模型处理通过拍摄特征和部件特征，以检测当前帧是否满足预先设定的拍摄规则，从而基于检测结果确定针对当前帧的视频拍摄引导策略。
根据一个实施例，当前帧是从车辆视频中按照预定时间间隔抽取的图像帧。
根据一个实施例，拍摄状态信息包括以下中的一项或多项：通过设于拍摄终端中的加速度传感器获取的加速度大小、加速度方向信息，通过陀螺仪获取的放置方向信息，通过定位模块获取的位置信息。
根据一个实施例，图像质量特征包括以下中的至少一项：图像是否清晰、图像是否为车辆图像、光线是否充足、车身是否有污渍。
根据一个实施例，有效性检测单元63还配置为：将图像质量特征和拍摄特征拼接后输入预先训练的第一检测模型，以检测当前帧的有效性；其中，当前帧的有效性还包括，是否仰拍图像或俯拍图像。
根据一个实施例，装置600还包括：图像拍摄引导单元(未示出)，配置为在检测到当前帧不是有效帧的情况下，提供针对当前帧的图像拍摄引导策略，图像拍摄引导策略包括以下中的一项：对准车辆拍摄、在光线充足时拍摄、清洗污渍后拍摄、保持拍摄终端沿竖直方向。
根据一个实施例，拍摄规则包括图像构图规则，图像构图规则指示预定部件落入图像中的预定区域，视频拍摄引导策略包括调整拍摄终端与车身的距离。
根据一个实施例，拍摄规则包括移动方向规则，用于检测拍摄终端的移动方向是否沿预定方向，视频拍摄引导策略包括向当前移动方向的相反方向移动，或者返回原点拍摄。
根据一个实施例，拍摄规则包括拍摄角度规则，用于检测当前帧是否跨越预定拍摄角度。
值得说明的是，以上对图6所示的辅助用户拍摄车辆视频的装置600，与图2示出的方法实施例相对应，图2对应的方法实施例中的相应描述也适用于图6所示的辅助用户拍摄车辆视频的装置，在此不再赘述。
根据另一方面的实施例，还提供一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算机中执行时，令计算机执行相应描述的方法。
根据再一方面的实施例，还提供一种计算设备，包括存储器和处理器，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现相应描述的方法。
本领域技术人员应该可以意识到，在上述一个或多个示例中，本说明书实施例所描述的功能可以用硬件、软件、固件或它们的任意组合来实现。当使用软件实现时，可以将这些功能存储在计算机可读介质中或者作为计算机可读介质上的一个或多个指令或代码进行传输。
以上所述的具体实施方式，对本说明书的技术构思的目的、技术方案和有益效果进行了进一步详细说明，所应理解的是，以上所述仅为本本说明书的技术构思的具体实施方式而已，并不用于限定本说明书的技术构思的保护范围，凡在本本说明书的技术构思的技术方案的基础之上，所做的任何修改、等同替换、改进等，均应包括在本本说明书的技术构思的保护范围之内。

Claims (21)

  1. 一种辅助用户拍摄车辆视频的方法，其中，所述方法包括：
    获取用户拍摄的车辆视频中的当前帧，以及用于拍摄所述车辆视频的拍摄终端的当前拍摄状态信息；
    利用预先训练的图像分类模型处理所述当前帧，从而获取所述当前帧的图像质量特征，并从所述拍摄状态信息提取拍摄特征；
    至少将所述图像质量特征输入预先训练的第一检测模型，以检测所述当前帧的有效性；
    在检测到所述当前帧是有效帧的情况下，利用预先训练的部件识别模型识别所述当前帧中的车辆部件，以基于识别结果确定所述当前帧的部件特征；
    利用预先训练的第二检测模型处理通过所述拍摄特征和所述部件特征，以检测所述当前帧是否满足预先设定的拍摄规则，从而基于检测结果确定针对所述当前帧的视频拍摄引导策略。
  2. 根据权利要求1所述的方法，其中，所述当前帧是从所述车辆视频中按照预定时间间隔抽取的图像帧。
  3. 根据权利要求1所述的方法，其中，所述拍摄状态信息包括以下中的一项或多项：所述拍摄终端的加速度大小、加速度方向信息、放置方向信息、位置信息。
  4. 根据权利要求1所述的方法，其中，所述图像质量特征包括以下中的至少一项：图像是否清晰、图像是否为车辆图像、光线是否充足、车身是否有污渍。
  5. 根据权利要求1所述的方法，其中，至少将所述图像质量特征输入预先训练的第一检测模型，以检测所述当前帧的有效性包括：
    将所述图像质量特征和所述拍摄特征拼接后输入预先训练的第一检测模型，以检测所述当前帧的有效性；
    其中，所述当前帧的有效性还包括，是否仰拍图像或俯拍图像。
  6. 根据权利要求1所述的方法，其中，所述方法还包括：
    在检测到所述当前帧不是有效帧的情况下，提供针对所述当前帧的图像拍摄引导策略，所述图像拍摄引导策略包括以下中的一项：对准车辆拍摄、在光线充足时拍摄、清洗污渍后拍摄、保持拍摄终端沿竖直方向。
  7. 根据权利要求1所述的方法，其中，所述拍摄规则包括图像构图规则，所述图像构图规则指示预定部件落入图像中的预定区域，所述视频拍摄引导策略包括调整所述拍摄终端与车身的距离。
  8. 根据权利要求1所述的方法，其中，所述拍摄规则包括移动方向规则，用于检测拍摄终端的移动方向是否沿预定方向，所述视频拍摄引导策略包括向当前移动方向的相反方向移动，或者返回原点拍摄。
  9. 根据权利要求1所述的方法，其中，所述拍摄规则包括拍摄角度规则，用于检测所述当前帧是否跨越预定拍摄角度。
  10. 根据权利要求1所述的方法，其中，所述第二检测模型是长短期记忆模型。
  11. 一种辅助用户拍摄车辆视频的装置，其中，所述装置包括：
    获取单元，配置为获取用户拍摄的车辆视频中的当前帧，以及用于拍摄所述车辆视频的拍摄终端的当前拍摄状态信息；
    第一特征提取单元，配置为利用预先训练的图像分类模型处理所述当前帧，从而获取所述当前帧的图像质量特征，并从所述拍摄状态信息提取拍摄特征；
    有效性检测单元，配置为至少将所述图像质量特征输入预先训练的第一检测模型，以检测所述当前帧的有效性；
    第二特征提取单元，配置为在检测到所述当前帧是有效帧的情况下，利用预先训练的部件识别模型识别所述当前帧中的车辆部件，以基于识别结果确定所述当前帧的部件特征；
    视频拍摄引导单元，配置为利用预先训练的第二检测模型处理通过所述拍摄特征和所述部件特征，以检测所述当前帧是否满足预先设定的拍摄规则，从而基于检测结果确定针对所述当前帧的视频拍摄引导策略。
  12. 根据权利要求11所述的装置，其中，所述当前帧是从所述车辆视频中按照预定时间间隔抽取的图像帧。
  13. 根据权利要求11所述的装置，其中，所述拍摄状态信息包括以下中的一项或多项：所述拍摄终端的加速度大小、加速度方向信息、放置方向信息、位置信息。
  14. 根据权利要求11所述的装置，其中，所述图像质量特征包括以下中的至少一项：图像是否清晰、图像是否为车辆图像、光线是否充足、车身是否有污渍。
  15. 根据权利要求11所述的装置，其中，所述有效性检测单元还配置为：
    将所述图像质量特征和所述拍摄特征拼接后输入预先训练的第一检测模型，以检测所述当前帧的有效性；
    其中，所述当前帧的有效性还包括，是否仰拍图像或俯拍图像。
  16. 根据权利要求11所述的装置，其中，所述装置还包括：
    图像拍摄引导单元，配置为在检测到所述当前帧不是有效帧的情况下，提供针对所 述当前帧的图像拍摄引导策略，所述图像拍摄引导策略包括以下中的一项：对准车辆拍摄、在光线充足时拍摄、清洗污渍后拍摄、保持拍摄终端沿竖直方向。
  17. 根据权利要求11所述的装置，其中，所述拍摄规则包括图像构图规则，所述图像构图规则指示预定部件落入图像中的预定区域，所述视频拍摄引导策略包括调整所述拍摄终端与车身的距离。
  18. 根据权利要求11所述的装置，其中，所述拍摄规则包括移动方向规则，用于检测拍摄终端的移动方向是否沿预定方向，所述视频拍摄引导策略包括向当前移动方向的相反方向移动，或者返回原点拍摄。
  19. 根据权利要求11所述的装置，其中，所述拍摄规则包括拍摄角度规则，用于检测所述当前帧是否跨越预定拍摄角度。
  20. 一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算机中执行时，令计算机执行权利要求1-10中任一项的所述的方法。
  21. 一种计算设备，包括存储器和处理器，其特征在于，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现权利要求1-10中任一项所述的方法。
PCT/CN2020/110735 2019-10-30 2020-08-24 辅助用户拍摄车辆视频的方法及装置 WO2021082662A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911046418.5 2019-10-30
CN201911046418.5A CN110650292B (zh) 2019-10-30 2019-10-30 辅助用户拍摄车辆视频的方法及装置

Publications (1)

Publication Number Publication Date
WO2021082662A1 true WO2021082662A1 (zh) 2021-05-06

Family

ID=68995126

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/110735 WO2021082662A1 (zh) 2019-10-30 2020-08-24 辅助用户拍摄车辆视频的方法及装置

Country Status (2)

Country Link
CN (2) CN113038018B (zh)
WO (1) WO2021082662A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114972298A (zh) * 2022-06-16 2022-08-30 中国电建集团中南勘测设计研究院有限公司 一种城市排水管道视频的检测方法及系统
CN115022671A (zh) * 2022-06-02 2022-09-06 智道网联科技(北京)有限公司 多进程视频输出方法、云端、终端、电子设备及存储介质
CN115107022A (zh) * 2022-06-16 2022-09-27 华中科技大学 一种基于gbdt算法的工业机器人位置误差补偿方法和系统

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113038018B (zh) * 2019-10-30 2022-06-28 支付宝(杭州)信息技术有限公司 辅助用户拍摄车辆视频的方法及装置
CN111652087B (zh) * 2020-05-15 2023-07-18 泰康保险集团股份有限公司 验车方法、装置、电子设备和存储介质
CN112492105B (zh) * 2020-11-26 2022-04-15 深源恒际科技有限公司 一种基于视频的车辆外观部件自助定损采集方法及系统
CN113408465B (zh) * 2021-06-30 2022-08-26 平安国际智慧城市科技股份有限公司 身份识别方法、装置及相关设备
CN114241180A (zh) * 2021-12-15 2022-03-25 平安科技（深圳）有限公司 车损理赔的图片检测方法、装置、计算机设备及存储介质
WO2023178589A1 (zh) * 2022-03-24 2023-09-28 深圳市大疆创新科技有限公司 拍摄指引方法、电子设备、系统及存储介质
CN115953726B (zh) * 2023-03-14 2024-02-27 深圳中集智能科技有限公司 机器视觉的集装箱箱面破损检测方法和系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018057497A1 (en) * 2016-09-21 2018-03-29 Allstate Insurance Company Enhanced image capture and analysis of damaged tangible objects
US10083487B1 (en) * 2013-03-13 2018-09-25 Liberty Mutual Insurance Company Imagery quantification of damage
CN109145903A (zh) * 2018-08-22 2019-01-04 阿里巴巴集团控股有限公司 一种图像处理方法和装置
CN110033386A (zh) * 2019-03-07 2019-07-19 阿里巴巴集团控股有限公司 车辆事故的鉴定方法及装置、电子设备
CN110650292A (zh) * 2019-10-30 2020-01-03 支付宝(杭州)信息技术有限公司 辅助用户拍摄车辆视频的方法及装置

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010032282A1 (ja) * 2008-09-16 2010-03-25 パイオニア株式会社 サーバ装置、移動端末装置、交差点案内システム及び交差点案内方法
CN104268783B (zh) * 2014-05-30 2018-10-26 翱特信息系统（中国）有限公司 车辆定损估价的方法、装置和终端设备
US20170291550A1 (en) * 2014-09-24 2017-10-12 Panasonic Intellectual Property Management Co., Ltd. On-board electronic mirror
CN107368776B (zh) * 2017-04-28 2020-07-03 阿里巴巴集团控股有限公司 车辆定损图像获取方法、装置、服务器和终端设备
CN107194323B (zh) * 2017-04-28 2020-07-03 阿里巴巴集团控股有限公司 车辆定损图像获取方法、装置、服务器和终端设备
CN108632530B (zh) * 2018-05-08 2021-02-23 创新先进技术有限公司 一种车辆定损的数据处理方法、装置、设备及客户端、电子设备
CN109325488A (zh) * 2018-08-31 2019-02-12 阿里巴巴集团控股有限公司 用于辅助车辆定损图像拍摄的方法、装置及设备
CN110245552B (zh) * 2019-04-29 2023-07-18 创新先进技术有限公司 车损图像拍摄的交互处理方法、装置、设备及客户端

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10083487B1 (en) * 2013-03-13 2018-09-25 Liberty Mutual Insurance Company Imagery quantification of damage
WO2018057497A1 (en) * 2016-09-21 2018-03-29 Allstate Insurance Company Enhanced image capture and analysis of damaged tangible objects
CN109145903A (zh) * 2018-08-22 2019-01-04 阿里巴巴集团控股有限公司 一种图像处理方法和装置
CN110033386A (zh) * 2019-03-07 2019-07-19 阿里巴巴集团控股有限公司 车辆事故的鉴定方法及装置、电子设备
CN110650292A (zh) * 2019-10-30 2020-01-03 支付宝(杭州)信息技术有限公司 辅助用户拍摄车辆视频的方法及装置

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115022671A (zh) * 2022-06-02 2022-09-06 智道网联科技(北京)有限公司 多进程视频输出方法、云端、终端、电子设备及存储介质
CN115022671B (zh) * 2022-06-02 2024-03-01 智道网联科技(北京)有限公司 多进程视频输出方法、云端、终端、电子设备及存储介质
CN114972298A (zh) * 2022-06-16 2022-08-30 中国电建集团中南勘测设计研究院有限公司 一种城市排水管道视频的检测方法及系统
CN115107022A (zh) * 2022-06-16 2022-09-27 华中科技大学 一种基于gbdt算法的工业机器人位置误差补偿方法和系统
CN114972298B (zh) * 2022-06-16 2024-04-09 中国电建集团中南勘测设计研究院有限公司 一种城市排水管道视频的检测方法及系统
CN115107022B (zh) * 2022-06-16 2024-04-19 华中科技大学 一种基于gbdt算法的工业机器人位置误差补偿方法和系统

Also Published As

Publication number Publication date
CN110650292A (zh) 2020-01-03
CN110650292B (zh) 2021-03-02
CN113038018A (zh) 2021-06-25
CN113038018B (zh) 2022-06-28

Similar Documents

Publication Publication Date Title
WO2021082662A1 (zh) 辅助用户拍摄车辆视频的方法及装置
TWI709091B (zh) 圖像處理方法和裝置
WO2020042800A1 (zh) 用于辅助车辆定损图像拍摄的方法、装置及设备
US10620005B2 (en) Building height calculation method, device, and storage medium
CN109495686B (zh) 拍摄方法及设备
US11024052B2 (en) Stereo camera and height acquisition method thereof and height acquisition system
CN108184048A (zh) 图像处理装置以及方法
CN111625091A (zh) 基于ar眼镜的标签叠加方法及装置
CN110427810B (zh) 视频定损方法、装置、拍摄端及机器可读存储介质
CN110263615A (zh) 车辆拍摄中的交互处理方法、装置、设备及客户端
CN111553302A (zh) 关键帧选取方法、装置、设备及计算机可读存储介质
CN113920502A (zh) 一种云台调整方法、装置、设备及介质
CN115965934A (zh) 一种停车位检测方法及装置
CN110658918B (zh) 用于视频眼镜的眼球追踪相机的定位方法、设备及介质
CN114693614A (zh) 防震锤缺陷检测方法、装置、设备及存储介质
CN111476056B (zh) 目标物体的识别方法、装置、终端设备及计算机存储介质
CN109460077B (zh) 一种自动跟踪方法、自动跟踪设备及自动跟踪系统
CN111402610A (zh) 一种识别红绿灯亮灯状态的方法、装置、设备及存储介质
US11875080B2 (en) Object sharing method and apparatus
CN113190013A (zh) 控制终端运动的方法和装置
CN115134533A (zh) 自动调用车载图像采集装置的拍摄方法及设备
CN113177967A (zh) 视频数据的物体追踪方法、系统和存储介质
CN113824874A (zh) 辅助摄像方法、装置、电子设备及存储介质
CN112565586A (zh) 一种自动对焦方法及装置
KR102628714B1 (ko) 모바일 단말용 사진 촬영 지원 카메라 시스템 및 방법

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: 20883249

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20883249

Country of ref document: EP

Kind code of ref document: A1