Method, device and equipment for analyzing collision test video

Abstract

The invention relates to the field of video analysis, and particularly discloses a collision test video analysis method, device and equipment, and the method comprises the steps: taking a moment when a collision signal is obtained as a starting moment, and determining a first to-be-processed video; excluding a non-interested area in the first to-be-processed video to obtain a second to-be-processed video; in response to the fact that all the first to-be-processed videos are converted into second to-be-processed videos, determining a moving track of each detection identifier in each video frame in the second to-be-processed videos and a matching relationship with each part of a target object; and determining a target video by taking a preset displacement moment of a preset part as an ending moment in the moving track corresponding to each detection identifier. According to the method, the starting moment and the ending moment of the collision test video are automatically identified, and the non-interested area is eliminated, so that the calculation amount required for analyzing the collision test video is reduced, and the analysis efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of video analysis, specifically to a method, apparatus, and equipment for analyzing collision test videos. Background Technology

[0002] In vehicle crash tests, such as Figure 1 As shown, in order to assess the displacement, acceleration and injury risk of volunteers or dummies in a collision, it is usually necessary to attach several optical markers to key parts of the human body (such as the head, shoulders, knees, etc.), place reference markers in non-deformable areas, take pictures with a high-speed camera, process the video after the test, establish a coordinate system, calculate the changes of each marker point in the coordinate system, and thus calculate the displacement, velocity, angle and other information of the marker point.

[0003] However, in the existing technology, when analyzing collision videos, it mainly relies on manually extracting collision video segments for analysis and identifying and analyzing the entire video frame. This not only requires excessive computing power but also has low recognition and analysis efficiency. Summary of the Invention

[0004] In view of the above problems, this disclosure provides a method, apparatus, and equipment for analyzing crash test videos to overcome or at least partially solve the above problems. The technical solution is as follows: A method for analyzing collision test videos includes: using the moment a collision signal is acquired as a start time to determine a first video to be processed; excluding non-interest regions in the first video to be processed to obtain a second video to be processed; in response to all first videos to be processed being converted into second videos to be processed, determining the movement trajectory of each detection marker within each video frame of the second video to be processed and its matching relationship with each part of the target object; and using a preset displacement moment of a preset part as an end time in the movement trajectory corresponding to each detection marker to determine the target video.

[0005] This application avoids the problems of wasted computational resources and tedious manual screening caused by the inability to automatically identify intervals in traditional analysis methods by limiting the motion analysis interval to "the time from the moment the collision signal is acquired to the preset displacement time of the preset location". Furthermore, it removes external interference by incorporating the elimination of non-interesting regions. Compared to traditional manual annotation or simple color detection algorithms, the technical solution proposed in this application has a higher degree of automation, accuracy, and efficiency, significantly reducing the false recognition rate and improving the overall quality of collision test video analysis.

[0006] In one example, the step of excluding non-interest regions in the first video to be processed to obtain a second video to be processed specifically includes: identifying each video frame of the first video to be processed to determine the region of interest and the location of all detection markers in each video frame; excluding non-interest regions based on the location of the detection markers and the region of interest, and determining the coordinates of each detection marker within the region of interest to obtain the second video to be processed.

[0007] In one example, determining the movement trajectory of each detection identifier within each video frame of the second video to be processed and its matching relationship with each part of the target object specifically includes: tracking the detection identifiers of each video frame in the second video to be processed to obtain the movement trajectory corresponding to each detection identifier; and determining the object part corresponding to each detection identifier based on a pre-constructed identifier topology template.

[0008] This application utilizes human detection technology to accurately identify human body regions in video frames and precisely locate the position of detection markers. This step effectively eliminates interference from non-human body regions and external detection markers, providing a clean data foundation for subsequent analysis. By excluding irrelevant regions, not only is the computational load reduced, but the accuracy and reliability of the analysis are also significantly improved.

[0009] In one example, the step of tracking the detection markers of each video frame in the second video to be processed to obtain the movement trajectory corresponding to each detection marker specifically includes: determining the detection boxes of all detection markers in each video frame, wherein the detection boxes are bounding boxes containing the detection markers; predicting the bounding boxes of each detection marker in the previous video frame in the next video frame as the prediction boxes of the next video frame; extracting the first appearance feature vector of the detection boxes and the second appearance feature vector of the prediction boxes; determining the intersection-union distance matrix between the detection boxes and the prediction boxes in the same video frame, and the cosine distance matrix between the first appearance feature vector and the second appearance feature vector; determining the cost matrix based on the intersection-union distance matrix and the cosine distance matrix; and determining the movement trajectory corresponding to each detection marker based on the cost matrix.

[0010] This application achieves accurate tracking of the movement trajectory of detected markers across video frames by comprehensively utilizing multiple techniques, including detection box localization, prediction box generation, appearance feature extraction, and distance matrix calculation. Specifically, firstly, the detection boxes of all detected markers in each video frame are determined; these boxes, as bounding boxes containing the detected markers, provide the foundation for subsequent tracking. Next, the bounding boxes of each detected marker in the previous video frame are predicted in the next video frame; this step helps to anticipate the possible positions of the detected markers. Then, the first appearance feature vector of the detection box and the second appearance feature vector of the prediction box are extracted. These feature vectors reflect the appearance characteristics of the detection and prediction boxes, providing a basis for subsequent matching and tracking. Based on this, the intersection-union (IU) distance matrix between the detection and prediction boxes in the same video frame, and the cosine distance matrix between the first and second appearance feature vectors, are determined. These two distance matrices measure the similarity between the detection and prediction boxes from both positional and appearance perspectives. Subsequently, based on the IU distance matrix and the cosine distance matrix, a cost matrix is ​​determined, which comprehensively reflects the cost or difficulty of moving the detected markers between different video frames. Finally, based on the cost matrix, the movement trajectory corresponding to each detection marker is determined. This step enables accurate tracking of the movement path of the detection marker between video frames.

[0011] In one example, determining the object part corresponding to each detection identifier based on a pre-built identifier topology template specifically includes: obtaining a pre-built identifier topology template, wherein the identifier topology template defines the relative positional relationship between each object part and the detection identifier; calculating the relative positional relationship between each detection identifier in the current video frame to form a detection position matrix; matching the detection position matrix with the relative positional relationship of the identifier topology template and determining the matching cost; and determining the matching result of the object part corresponding to each detection identifier based on the matching cost.

[0012] This application clarifies the relative positional relationships between each object part and the detected markers through a pre-constructed marker topology template, providing a benchmark for the subsequent matching process. In practice, the relative positional relationships between each detected marker in the current video frame are first calculated to form a detection position matrix, which reflects the spatial distribution of the detected markers in the current frame. Next, the detection position matrix is ​​matched with the relative positional relationships in the marker topology template, and the matching cost is determined by comparing the similarity between the two. The matching cost reflects the degree of agreement between the detected marker positions and the relative positional relationships defined by the topology template. Finally, based on the matching cost, the matching result for each object part corresponding to the detected marker is determined. This step enables automatic identification of the human body parts to which the detected markers belong, improving the accuracy and reliability of the identification. By introducing the marker topology template, this application can address the problem of detecting marker position recognition under different object poses and viewing angle changes, providing more accurate and comprehensive information support for subsequent collision test data analysis.

[0013] In one example, determining the object location corresponding to each detected marker based on a pre-built marker topology template further includes: performing an affine transformation on the marker topology template to determine the position of the detected marker under different object poses and viewpoint changes; determining the confidence level of the matching result by the matching degree between the affine transformed marker topology template and the position of the detected marker in the current video frame; and when the confidence level is lower than a preset confidence threshold, re-performing topology matching or prompting manual intervention.

[0014] This application further considers the impact of changes in object posture and viewing angle on the recognition of detected marker positions. By introducing affine transformation technology, the topological template of the marker is dynamically adjusted to adapt to the recognition requirements of detected marker positions under different postures and viewing angles. Simultaneously, by calculating the confidence level of the matching results and setting a preset confidence threshold, when the confidence level falls below this threshold, an automatic re-matching or manual intervention mechanism is triggered, ensuring the accuracy and reliability of the recognition results. This technical approach effectively improves the automation and intelligence level of collision test video analysis.

[0015] In one example, after determining the target video by using a preset displacement time of a preset part as the end time in the movement trajectory corresponding to each detection marker, the method further includes: determining the movement data of the detection marker based on the movement trajectory of each detection marker, the movement data including at least one of displacement, velocity, and acceleration; generating displacement curves of candidate parts based on the movement trajectory of each detection marker; when the region of interest is a human body region, the candidate parts include at least one of head, neck, shoulder, elbow, wrist, spine, hip, knee, and ankle; extracting the peak value of a preset type of motion index, the preset type of motion index including at least one of maximum displacement, peak velocity, and acceleration extreme value; and generating an analysis report based on the movement data, the peak value of the preset type of motion index, and the displacement curves of the candidate parts.

[0016] This application, by deeply mining the movement trajectory data of each detection marker, not only calculates key movement data such as displacement, velocity, and acceleration of various parts of the object to be identified, but also provides a strong basis for analyzing the impact of the collision by visually demonstrating the dynamic changes of each part during the collision. Furthermore, by extracting the peak values ​​of preset types of motion indicators such as maximum displacement, peak velocity, and acceleration extremes, the severity of the collision and its impact on various parts of the object are further quantified. Finally, based on this comprehensive and accurate data, the automatically generated analysis report provides researchers conducting collision experiments with detailed and intuitive data support, helping them to deeply understand the motion characteristics and force conditions of the object during the collision. In addition, the automated report generation process greatly improves work efficiency, reduces the time cost of manual data processing and analysis, and allows researchers to focus more on the interpretation and application of experimental results.

[0017] In one example, after determining the first video to be processed by taking the moment when the collision signal is obtained as the starting moment, the method further includes: obtaining the camera-involved radial distortion coefficients corresponding to the initial collision video; and performing distortion model correction on the first video to be processed based on the camera-involved radial distortion coefficients.

[0018] This application, after obtaining the radial distortion coefficients of the camera's internal components corresponding to the initial collision video, constructs an accurate distortion model using these parameters. This model accurately describes the distortion characteristics generated by the camera lens during imaging, including different types of distortion such as radial distortion. Subsequently, this distortion model is applied to each frame of the first video to be processed, and the position of each pixel in the image is adjusted through inverse mapping or correction algorithms to eliminate or significantly reduce the impact of lens distortion. This step ensures that the video data on which subsequent analysis is based has higher geometric accuracy and visual realism, thereby improving the accuracy and reliability of the entire collision test video analysis process. Through distortion correction, measurement errors caused by lens distortion can be avoided, making the calculation of key parameters such as displacement, velocity, and acceleration based on video data more accurate.

[0019] This application also provides an analysis device for collision test videos, comprising: a first video to be processed determination module, which determines the first video to be processed by taking the moment when the collision signal is obtained as the start time; a region exclusion module, which excludes non-interest regions in the first video to be processed to obtain a second video to be processed; a detection marker tracking module, which determines the movement trajectory of each detection marker in each video frame of the second video to be processed and its matching relationship with each part of the target object in response to all the first videos to be processed being converted into the second videos to be processed; and a target video determination module, which determines the target video by taking the preset displacement time of a preset part as the end time in the movement trajectory corresponding to each detection marker.

[0020] This application also provides an analysis device for crash test videos, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: implement the crash test video analysis method as described in any of the above examples.

[0021] By employing the aforementioned technical solutions, this disclosure provides a method, apparatus, and device for analyzing collision test videos. By automatically identifying the start and end times of the collision test video and eliminating non-interesting regions, it reduces the computational load required for analyzing the video, significantly improving the automation, accuracy, and efficiency of collision test video analysis. Compared to traditional manual annotation or simple color detection algorithms, this disclosure's technical solution can automatically identify key analysis intervals, reducing wasted computational resources and the tedious process of manual screening. Simultaneously, by eliminating non-interesting regions and interference from external detection markers, it provides a clean data foundation for subsequent analysis, significantly improving the accuracy and reliability of the analysis. Furthermore, automatically tracking the movement trajectory of the detection markers and generating detailed analysis reports provides comprehensive and accurate data support for collision test data analysis, helping researchers to deeply understand the motion characteristics and force conditions of objects during the collision process.

[0022] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, specific embodiments of this disclosure are described below. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this disclosure. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A schematic diagram of a dummy and seat with optical markings affixed in an embodiment of this disclosure is shown; Figure 2 A flowchart illustrating a method for analyzing a collision test video according to an embodiment of this disclosure is shown; Figure 3 This diagram illustrates the displacement response curve of a candidate site according to an embodiment of the present disclosure. Figure 4 A schematic diagram of the structure of a collision test video analysis device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0025] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0026] In vehicle crash tests, such as Figure 1 As shown, to assess the displacement, acceleration, and injury risk of volunteers or dummies in a collision, it is usually necessary to attach several optical markers to key parts of the object to be analyzed (such as the human body) (such as the head, shoulders, knees, etc.), place reference markers in non-deformable areas, take pictures with a high-speed camera, process the video after the test, establish a coordinate system, calculate the changes of each marker point in the coordinate system, and thus calculate the displacement, velocity, angle and other information of the marker point.

[0027] In existing technologies, marker recognition and tracking mainly rely on manual annotation or simple color detection algorithms, which are not only inefficient and have a high false recognition rate, but also struggle to handle issues such as multiple markers with the same shape, changes in lighting, and distortion interference. Furthermore, test videos are typically long; for example, the meaningful motion analysis interval for a human is only between the moment the collision signal is acquired and the head rebound frame after braking. Traditional analysis methods cannot automatically identify this interval, leading to wasted computational resources and tedious manual screening. Simultaneously, since vehicle-mounted cameras often use wide-angle lenses, the images exhibit significant distortion, further affecting the accuracy of marker detection. Therefore, there is an urgent need for a low-false-recognition-rate tracking system capable of automatic marker recognition without manual annotation, capable of automatically extracting the effective analysis interval and automatically correcting distortion, to improve the automation level and analysis efficiency of crash tests.

[0028] Therefore, this application provides a method, apparatus, and equipment for analyzing collision test videos, such as... Figure 1 The diagram illustrates a flowchart of a method for analyzing crash test videos provided in one or more embodiments of this specification. This method can be applied to analyze vehicle crash test videos. The process can be executed by a computing device used for video analysis (e.g., a control terminal connected to a camera, or a server deployed in the cloud). Certain input parameters or intermediate results in the process can be manually adjusted to help improve accuracy.

[0029] The analysis method involved in the embodiments of this application can be implemented by a terminal device or a server, and this application does not impose any special limitations on it. For ease of understanding and description, the following embodiments are all described in detail using a server as an example.

[0030] It should be noted that the server can be a single device or a system composed of multiple devices, i.e., a distributed server. This application does not make any specific limitations in this regard.

[0031] like Figure 2 As shown in the embodiment of this application, a method for analyzing collision test videos is provided, including: S201: Use the moment when the collision signal is obtained as the starting moment to determine the first video to be processed.

[0032] First, after the server acquires the initial collision video, it uses the moment when the collision signal is detected (e.g., the moment the flash is triggered) as the key time node. Starting from this point, subsequent video segments are extracted to determine the first video to be processed. The significance of this step lies in precisely locking down the starting range for subsequent analysis. Because the moment the collision signal is detected marks the official start of the collision test, only the video content from this point onward contains the collision process information we need to analyze. This avoids processing irrelevant video content from before the test, improving analysis efficiency and focus.

[0033] S202: Remove the non-interested regions from the first video to be processed to obtain the second video to be processed.

[0034] After determining the first video to be processed, in actual crash test videos, besides the region of interest (ROI) and its attached optical markers, there are other interfering elements, such as fixed objects in the test environment and other irrelevant markers. These ROIs and external detection markers can interfere with subsequent marker tracking and object part matching operations, affecting the accuracy of the analysis results. Therefore, by excluding ROIs and external detection markers from the first video to be processed, a second video to be processed can be obtained, further reducing the computational load and focusing subsequent processing on data truly relevant to the crash test. Here, external detection markers refer to those located within ROIs, while ROIs refer to the areas containing key parts of the object being analyzed (such as the human body), which are crucial for the analysis of crash test results. By focusing on ROIs, the accuracy of marker tracking can be effectively improved, and interference from irrelevant data can be reduced. For example, in crash tests involving dummy models, ROIs typically include key parts such as the head, neck, and shoulders, whose motion trajectories and stress conditions directly affect the assessment of crash safety performance. Correspondingly, regions of non-interest refer to parts of the test environment that are irrelevant to collision analysis, such as background walls, fixed equipment, or other non-critical objects. These regions not only do not contribute to the evaluation of collision results but may also introduce noise and interference, affecting the recognition accuracy and tracking stability of detection markers. Therefore, automatically identifying and excluding these regions of non-interest through intelligent algorithms can significantly improve the efficiency and accuracy of video analysis.

[0035] S203: In response to all first videos to be processed being converted into second videos to be processed, determine the movement trajectory of each detection marker in each video frame of the second video to be processed and its matching relationship with each part of the target object.

[0036] After all the first-stage videos are converted into second-stage videos, the server tracks the detection markers in each frame of the second-stage video to obtain the movement trajectory corresponding to each marker. In the collision test, the optical markers on key parts of the target object move with the object's motion. By tracking the positional changes of these markers in different video frames, their movement trajectories can be obtained. These trajectories contain motion information of various parts of the object during the collision process, such as displacement and velocity. After obtaining the movement trajectory corresponding to each marker, the server determines the object part corresponding to each marker based on a pre-built marker topology template. This marker topology template is pre-set according to the structural characteristics of the target object and the marker's placement in the collision test; it accurately maps the correspondence between each marker and the object part. In this way, the server can automatically identify the specific object part represented by each marker, providing an accurate data foundation for subsequent analysis.

[0037] For example, if a marker is identified as corresponding to an object's head, the server knows that the marker's movement trajectory actually reflects the head's movement. Automating this step not only improves analysis efficiency but also reduces errors that might arise from manual annotation.

[0038] S204: In the movement trajectory corresponding to each detection marker, the preset displacement time of the preset part is used as the end time to determine the target video.

[0039] After identifying the corresponding human body parts for each detection marker, the server needs to further determine the target video range for analysis. Since the truly meaningful motion analysis interval in a crash test is usually within a specific range—for example, for a dummy, this specific range is between the "flash trigger frame and the head rebound frame after braking"—the server uses the preset displacement time of a preset part of the object as the end time and extracts a video segment within this critical range from the second video to be processed as the target video. The precise execution of this step ensures that subsequent analysis can focus on the most critical stage of the object's motion during the collision, thereby improving the relevance and accuracy of the analysis.

[0040] For ease of explanation, the region of interest will be described using the human body as an example. In this case, the target object is the dummy's torso, the preset part is the dummy's head, and the preset displacement time is the head's rebound time.

[0041] In one embodiment, when excluding non-human regions of interest (MOIs) and external detection markers in the first video to be processed, it is first necessary to identify each video frame of the first video to determine the MOI (human body region) and the location of the detection markers in each video frame. Then, based on the location of the detection markers and the human body region, non-human regions can be excluded to obtain the second video to be processed. It should be noted that although only non-human regions are excluded here, the process of identifying detection markers is to determine the location of each detection marker. The detection markers outside the human body are reference markers used to estimate the world coordinates of other measurement markers. They are almost stationary, and generally only one video frame is needed to detect them, without further tracking. The remaining detection markers are measurement markers attached to the human body. Each frame needs to be identified and linked into a trajectory using a tracking algorithm. The purpose of human body detection is to eliminate interference from misidentification of markers outside the human body.

[0042] In the detection and recognition of identifiers and human bodies, object detection algorithms can be used. These algorithms can be broadly categorized into three types: feature-based methods, two-stage detectors based on deep learning, and single-stage detectors based on deep learning. A representative feature-based algorithm is the Histogram of Oriented Gradients (HOG) combined with Support Vector Machines (SVM). Different types of algorithms are suitable for different scenarios, with single-stage detectors being suitable for scenarios requiring high processing speed.

[0043] The source data for object detection here is a sequence of video frames, which is large in volume. Therefore, a single-stage detector, such as the YOLO11 object detection algorithm, is more suitable. Compared with single-stage detectors, YOLO11 has higher accuracy, achieves a good balance between speed and accuracy, converges faster during training, and has lower resource requirements.

[0044] Taking human detection model training as an example, the model training process is as follows: Video frames contain a large amount of redundant information; keyframes are extracted to remove redundant frames. There are approximately 200 images in total, divided into training, validation, and test sets in an 8:1:1 ratio. The acquired video data is labeled; for example, the computer vision annotation tool CVAT can be used to annotate measurement markers (black and yellow alternating circles). During model training, the model trained on a large visualization database (such as ImageNet) used for visual object recognition software research is fine-tuned using the above data. The training data is augmented, including adjusting brightness, contrast, mirroring, and perspective changes; hyperparameters such as the number of training epochs, learning rate, and optimizer are also adjusted.

[0045] When performing target tracking, the detected markers can be tracked using target tracking algorithms. These algorithms can be divided into single-target tracking and multi-target tracking methods. Multi-target tracking has many advantages over single-target tracking, including but not limited to: strong versatility, no need for manual specification, high degree of automation, stronger robustness, and the ability to maintain the correct identity of the target in complex scenarios.

[0046] Therefore, this application can track detected markers using the BoT-SORT multi-object tracking algorithm: determining the detection boxes of all detected markers in each video frame, wherein the detection boxes are bounding boxes containing the detected markers; predicting the bounding boxes of each detected marker in the previous video frame in the next video frame, as the predicted boxes for the next video frame; extracting the first appearance feature vector of the detection box and the second appearance feature vector of the predicted box; determining the intersection-union distance matrix between the detection box and the predicted box in the same video frame, and the cosine distance matrix between the first appearance feature vector and the second appearance feature vector; determining the cost matrix based on the intersection-union distance matrix and the cosine distance matrix; and determining the movement trajectory corresponding to each detected marker based on the cost matrix.

[0047] Specifically, during object detection, for each input frame, the object detector first obtains the bounding boxes of all objects in the image, typically including the center coordinates, width, height, and confidence score of the boxes. For state prediction, an improved Kalman filter is used to predict the position of the currently tracked object in the current frame. Unlike traditional methods (such as Deep SORT) that estimate the aspect ratio of the bounding boxes, BoT-SORT's Kalman filter state vector directly estimates the width and height of the bounding boxes and uses time-dependent noise covariance, resulting in more accurate predicted bounding boxes.

[0048] Furthermore, to address the prediction box drift caused by camera motion, this application introduces a camera motion compensation mechanism. Background motion is compensated by tracking image keypoints (e.g., using the pyramid Lucas-Kanade optical flow method) and estimating the inter-frame affine transformation matrix using the RANSAC algorithm. The estimated affine transformation matrix is ​​then used to correct the Kalman filter's predictions, thereby stabilizing the target trajectory.

[0049] During appearance feature extraction, a pre-trained re-identification network is used to extract the appearance feature vector for each detection box. This feature vector is used to measure the appearance similarity between targets, helping to address tracking loss issues caused by occlusion and rapid movement.

[0050] To associate the current detection bounding box with the predicted tracking trajectory, a matching strategy that fuses motion and appearance information can be adopted during data association, considering both motion similarity and appearance similarity. Motion similarity is calculated by determining the intersection-union ratio (IU / R) distance matrix between the predicted and detection bounding boxes. Appearance similarity is calculated by determining the cosine distance matrix between the appearance features of the predicted trajectory and the appearance features of the detection bounding boxes.

[0051] During fusion matching, the intersection-over-union distance matrix and the cosine distance matrix can be fused (e.g., by taking the minimum of corresponding elements) to form a comprehensive cost matrix. Subsequently, the Hungarian algorithm is used for matching to correlate detection with trajectories.

[0052] During trajectory management, for successfully matched trajectories, their position, velocity (via Kalman filtering), and appearance status (e.g., using exponential moving average (EMA) to update appearance features) can be updated using the corresponding detection boxes. Unmatched detections may be initialized as new trajectories. Unmatched trajectories are marked as "missing." If a trajectory remains missing for more than a certain number of frames, it will be terminated and removed. By applying a multi-target tracking algorithm, automatic and accurate tracking of multiple detection markers can be achieved. In a crash test video, multiple optical markers are attached to a human body. These markers move with the human's movement. The multi-target tracking algorithm can simultaneously track the positional changes of these markers in different video frames, obtaining their individual movement trajectories. In this process, the algorithm first determines the bounding boxes of all detection markers in each video frame, accurately defining their positions. Then, the algorithm predicts the bounding boxes of each detection marker in the previous video frame for the next video frame, serving as the predicted bounding boxes. Next, the algorithm extracts the appearance feature vectors of the detection and predicted bounding boxes, and determines the cost matrix by calculating their intersection-over-union (IoU) and cosine distance matrices. Finally, based on the cost matrix, the algorithm accurately determines the movement trajectory of each detection marker. This multi-target tracking algorithm not only improves the accuracy and efficiency of tracking but also reduces errors that may arise from human intervention and subjective judgment, providing strong support for the automated analysis of crash test videos. Furthermore, the algorithm has strong versatility and robustness, adapting to the needs of crash test video analysis in different scenarios.

[0053] Compared with other similar algorithms (SORT, Deep SORT, etc.), BoT-SORT uses a more accurate Kalman filter state vector and introduces camera motion compensation technology. It combines the advantages of motion information and appearance information, and can establish a stronger correlation between detection and trajectory, thus maintaining a high level of identity preservation while ensuring high tracking accuracy.

[0054] In one embodiment, when determining the object part corresponding to each detection identifier, a detection position matrix can be formed by calculating the relative positional relationship between each detection identifier in the current video frame; then, the detection position matrix is ​​matched with the relative positional relationship of the identifier topology template, and the matching cost is determined; finally, the matching result of the object part corresponding to each detection identifier is determined based on the matching cost.

[0055] This application, through the aforementioned matching method, can automatically and accurately determine the object parts corresponding to each detection marker. Specifically, in each frame of the collision test video, the relative positional relationships between the detection markers are fixed, and a marker topology template is pre-set based on the object structure and the marker's placement. A detection position matrix is ​​formed by calculating the relative positional relationships between the detection markers in the current video frame, and then meticulously matched with the relative positional relationships of the marker topology template, calculating the matching cost. The lower the matching cost, the higher the degree of matching between the current detection position matrix and the marker topology template. Based on this matching cost, the matching results of the object parts corresponding to each detection marker can be accurately obtained. This automated matching method greatly improves analysis efficiency, avoids errors that may occur with manual annotation, ensures that subsequent analysis is based on accurate data, and provides a reliable basis for a deeper understanding of the object's motion characteristics and force conditions during the collision process. Moreover, this method has strong adaptability and stability, and can accurately complete the matching of detection markers and object parts even in collision test videos with different scenarios and object postures.

[0056] Furthermore, after obtaining the identifier topology template, since the dummy may present different postures in the collision test and the corresponding shooting postures of the camera are different, the identifier topology template can be subjected to an affine transformation to determine the detection identifier position under different object postures and viewpoint changes; then, the confidence level of the matching result is determined by the matching degree between the identifier topology template after affine transformation and the detection identifier position in the current video frame; when the confidence level is lower than the preset confidence threshold, topology matching is performed again or manual intervention is prompted.

[0057] This application effectively improves the accuracy and reliability of matching detection markers with object parts by performing affine transformation on the marker topology template and subsequent related operations. In actual collision test scenarios, the poses of objects (such as dummies) are diverse, and the camera's shooting angles also vary, which undoubtedly poses a challenge to accurately matching detection markers with object parts. However, by affine transforming the marker topology template, the positional situation of the detection marker under different object poses and viewing angle changes can be simulated. Then, the matching degree between the affine transformed marker topology template and the actual position of the detection marker in the current video frame is calculated to determine the confidence level of the matching result. When the confidence level is lower than the preset confidence threshold, it indicates that the current automatic matching result may have a large error. At this time, it is necessary to re-perform topology matching or prompt manual intervention.

[0058] Re-performing topology matching allows for further attempts to find more suitable matching relationships using algorithms, while manual intervention leverages the experience and judgment of professionals to correct complex situations or those that algorithms struggle to handle accurately. This combined approach of automated processing and manual intervention ensures accurate identification of the object parts corresponding to each detection marker in various complex scenarios. This lays a solid foundation for generating accurate analysis reports and gaining a deeper understanding of the object's motion characteristics and force conditions during the collision, further improving the practicality and accuracy of the entire collision test video analysis method.

[0059] Finally, the server generates an analysis report of the target video based on the object parts corresponding to each detection marker and their movement trajectories. This report records in detail the motion of each object part during the collision, including key parameters such as displacement, velocity, and acceleration, as well as the trends of these parameters over time. Through this data, researchers can gain a deeper understanding of the motion characteristics and forces acting on objects during the collision, providing strong data support for vehicle safety design and crash test evaluation. Furthermore, because the entire analysis process is automated, it significantly improves efficiency and accuracy, reducing errors that may arise from human intervention and subjective judgment.

[0060] In one embodiment, when generating an analysis report, the movement data of each detection marker can be determined based on its movement trajectory. This movement data includes displacement, velocity, and acceleration. Based on the movement trajectories of each detection marker, displacement curves for candidate body parts are generated. These candidate body parts can include the head, neck, shoulder, elbow, wrist, spine, hip, knee, and ankle. Preset type motion index peak values ​​are extracted. These preset type motion indexes include maximum displacement, peak velocity, and acceleration extreme values. Based on the movement data, the preset type motion index peak values, and the displacement curves of the candidate body parts, the analysis report is generated. The displacement response curve of the candidate body part is shown below. Figure 3 As shown, Figure 3 The horizontal and vertical coordinates represent the coordinates of the measurement point on the XZ plane.

[0061] This application automatically generates analysis reports, comprehensively and accurately presenting key information from collision test videos. This significantly saves time and effort compared to manual analysis, improves efficiency, and avoids potential oversights and errors, ensuring the accuracy and reliability of the results. Specifically, determining movement data based on the trajectory of each detection marker provides fundamental quantitative information for subsequent analysis. Displacement, velocity, and acceleration data intuitively reflect the motion state of different parts of the object during the collision. Generating displacement curves for candidate parts graphically displays the motion of different parts of the object, allowing users to clearly observe the trends and patterns of motion. The displacement of key parts such as the head and neck during the collision is readily apparent. Extracting preset motion index peaks, such as maximum displacement, peak velocity, and acceleration extremes, highlights key characteristics of the object's motion during the collision, helping to quickly locate critical nodes that significantly impact the object.

[0062] In one embodiment, since vehicle-mounted cameras often use wide-angle lenses, the images exhibit significant distortion, affecting the accuracy of marker detection. Therefore, distortion correction can be performed on the first video to be processed after its initial determination. Specifically, the intrinsic parameters and distortion parameters of the vehicle-mounted camera can be obtained first. These parameters are typically calibrated at the camera's factory or recalibrated using specialized calibration tools during subsequent use. Using these parameters, a suitable distortion correction algorithm, such as a correction algorithm based on a radial distortion model, is employed to correct the distortion of each frame of the first video to be processed. The corrected video image effectively eliminates the distortion caused by the wide-angle lens, making the markers in the image more regular, thereby improving the accuracy of subsequent target detection and tracking operations. After distortion correction is completed, the subsequent analysis and processing are carried out according to the process described above, including target detection, target tracking, matching of detected markers with object parts, and generating an analysis report. This ensures that the entire collision test video analysis process is based on accurate video data, ultimately yielding reliable and accurate analysis results.

[0063] By employing the aforementioned technical solutions, this disclosure provides a method, apparatus, and equipment for analyzing collision test videos. Through a series of technical means, including precisely locking the motion analysis interval, effectively eliminating external interference, automatically tracking the movement trajectory of detection markers, and automatically generating detailed analysis reports, it significantly improves the automation, accuracy, and efficiency of collision test video analysis. Compared to traditional manual annotation or simple color detection algorithms, the technical solution of this disclosure can automatically identify key analysis intervals, reducing the waste of computational resources and the tedious process of manual screening. Simultaneously, by using object detection technology to eliminate interference from non-object areas and external detection markers, it provides a clean data foundation for subsequent analysis, significantly improving the accuracy and reliability of the analysis. Furthermore, automatically tracking the movement trajectory of detection markers and generating detailed analysis reports provides comprehensive and accurate data support for the data analysis of collision tests, helping researchers to deeply understand the motion characteristics and force conditions of objects during the collision process. In addition, the technical solution of this disclosure also possesses high flexibility and scalability, adapting to the analysis needs of different types of collision test videos. As long as identifiable detection markers exist in the video, this technical solution can leverage its advantages to perform accurate and effective analysis.

[0064] Figure 4 This is a schematic diagram of the structure of a collision test video analysis device provided in an embodiment of this application. The device includes: The first video to be processed determination module 401 uses the moment when the collision signal is obtained as the start time to determine the first video to be processed. The region exclusion module 402 excludes non-interest regions from the first video to be processed to obtain the second video to be processed; The detection marker tracking module 403, in response to all first videos to be processed being converted into second videos to be processed, determines the movement trajectory of each detection marker in each video frame of the second video to be processed and its matching relationship with each part of the target object. The target video determination module 404 determines the target video by taking the preset displacement time of the preset part as the end time in the movement trajectory corresponding to each detection mark.

[0065] In one specific embodiment, the region exclusion module 402 includes: identifying each video frame of the first video to be processed to determine the region of interest and the location of all detection markers in each video frame; excluding non-regions of interest based on the location of the detection markers and the region of interest, and determining the coordinates of each detection marker within the region of interest to obtain the second video to be processed.

[0066] In one specific embodiment, the detection identifier tracking module 403 includes: tracking the detection identifiers of each video frame in the second video to be processed to obtain the movement trajectory corresponding to each detection identifier; and determining the object part corresponding to each detection identifier based on a pre-constructed identifier topology template.

[0067] In one specific embodiment, the detection identifier tracking module 403 includes: determining detection boxes for all detection identifiers in each video frame, wherein the detection boxes are bounding boxes containing the detection identifiers; predicting the bounding boxes of each detection identifier in the previous video frame in the next video frame as the predicted boxes for the next video frame; extracting a first appearance feature vector of the detection boxes and a second appearance feature vector of the predicted boxes; determining the intersection-union distance matrix between the detection boxes and the predicted boxes in the same video frame, and the cosine distance matrix between the first appearance feature vector and the second appearance feature vector; determining a cost matrix based on the intersection-union distance matrix and the cosine distance matrix; and determining the movement trajectory corresponding to each detection identifier based on the cost matrix.

[0068] In one specific embodiment, the detection marker tracking module 403 includes: acquiring a pre-constructed marker topology template, wherein the marker topology template defines the relative positional relationship between each object part and the detection marker; calculating the relative positional relationship between each detection marker in the current video frame to form a detection position matrix; matching the detection position matrix with the relative positional relationship of the marker topology template and determining the matching cost; and determining the object part matching result corresponding to each detection marker based on the matching cost.

[0069] In one specific embodiment, the detection marker tracking module 403 includes: performing an affine transformation on the marker topology template to determine the position of the detected marker under different object poses and viewpoint changes; determining the confidence level of the matching result by the matching degree between the affine transformed marker topology template and the detected marker position in the current video frame; and when the confidence level is lower than a preset confidence threshold, re-performing topology matching or prompting manual intervention.

[0070] In one specific embodiment, the target video determination module 404 includes: determining the movement data of each detection marker based on the movement trajectory of each detection marker, wherein the movement data includes at least one of displacement, velocity, and acceleration; generating displacement curves of candidate parts based on the movement trajectories of each detection marker; when the region of interest is a human body region, the candidate parts include at least one of head, neck, shoulder, elbow, wrist, spine, hip, knee, and ankle; extracting the peak value of a preset type of motion index, wherein the preset type of motion index includes at least one of maximum displacement, peak velocity, and acceleration extreme value; and generating an analysis report based on the movement data, the peak value of the preset type of motion index, and the displacement curves of the candidate parts.

[0071] In one specific embodiment, the first video to be processed determination module 401 includes: obtaining the camera-involved radial distortion coefficients corresponding to the initial collision video; and performing distortion model correction on the first video to be processed based on the camera-involved radial distortion coefficients.

[0072] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0073] This application embodiment also provides an analysis device for collision test videos, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: determine a first video to be processed using the moment of obtaining a collision signal as a start time; exclude non-object regions and external detection markers in the first video to be processed to obtain a second video to be processed; in response to all first videos to be processed being converted into second videos to be processed, track the detection markers of each video frame in the second video to obtain the movement trajectory corresponding to each detection marker; determine the object part corresponding to each detection marker based on a pre-constructed marker topology template; determine a target video using the head rebound moment as the end time in the movement trajectory corresponding to each detection marker; and generate an analysis report of the target video based on the object part corresponding to each detection marker and the movement trajectory.

[0074] This application embodiment also provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: The moment when the collision signal is obtained is taken as the starting moment to determine the first video to be processed; the non-interest regions in the first video to be processed are excluded to obtain the second video to be processed; in response to all the first videos to be processed being converted into the second videos to be processed, the movement trajectory of each detection marker in each video frame of the second video to be processed and the matching relationship with each part of the target object are determined; in the movement trajectory corresponding to each detection marker, the preset displacement moment of the preset part is taken as the ending moment to determine the target video.

[0075] This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned steps to implement the collision test video analysis method provided in the above embodiment.

[0076] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0077] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0078] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0083] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0084] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0085] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for analyzing collision test videos, characterized in that, include: The moment when the collision signal is obtained is used as the starting moment to determine the first video to be processed; The non-interested regions in the first video to be processed are excluded to obtain the second video to be processed; In response to all first videos to be processed being converted into second videos to be processed, the movement trajectory of each detection marker in each video frame of the second video to be processed and its matching relationship with each part of the target object are determined. In the movement trajectory corresponding to each detection marker, the preset displacement time of the preset part is used as the end time to determine the target video.

2. The method according to claim 1, characterized in that, The step of excluding non-interesting regions from the first video to be processed to obtain the second video to be processed specifically includes: Each video frame of the first video to be processed is identified to determine the region of interest in each video frame and the location of all detection markers; Based on the location of the detection markers and the region of interest, the non-region of interest is excluded, and the coordinates of each detection marker within the region of interest are determined to obtain the second video to be processed.

3. The method according to claim 1, characterized in that, Determining the movement trajectory of each detection marker within each video frame of the second video to be processed and its matching relationship with each part of the target object specifically includes: The detection markers of each video frame in the second video to be processed are tracked to obtain the movement trajectory corresponding to each detection marker; Based on a pre-built identifier topology template, the object part corresponding to each detection identifier is determined.

4. The method according to claim 3, characterized in that, The step of tracking the detection markers of each video frame in the second video to be processed to obtain the movement trajectory corresponding to each detection marker specifically includes: Determine the detection bounding boxes for all detection markers in each video frame, wherein the detection bounding boxes are bounding boxes containing the detection markers; Predict the bounding boxes of each detection marker in the previous video frame in the next video frame, and use them as the predicted bounding boxes for the next video frame. Extract the first appearance feature vector of the detection box and the second appearance feature vector of the prediction box; Determine the intersection-union distance matrix between the detection box and the prediction box in the same video frame, and the cosine distance matrix between the first appearance feature vector and the second appearance feature vector; The cost matrix is ​​determined based on the intersection-union ratio distance matrix and the cosine distance matrix; Based on the cost matrix, the movement trajectory corresponding to each detection marker is determined.

5. The method according to claim 3, characterized in that, The process of determining the object location corresponding to each detected marker based on a pre-built marker topology template specifically includes: Obtain a pre-constructed identifier topology template, wherein the identifier topology template defines the relative positional relationship between each object part and the detection identifier; Calculate the relative positional relationships between each detection marker in the current video frame to form a detection position matrix; The relative positional relationship between the detected position matrix and the identifier topology template is matched, and the matching cost is determined; Based on the matching cost, the matching result of the object part corresponding to each detection mark is determined.

6. The method according to claim 5, characterized in that, The method of determining the object part corresponding to each detection marker based on a pre-built marker topology template also includes: An affine transformation is performed on the topological template of the marker to determine the position of the detected marker under different object poses and viewpoint changes; The confidence level of the matching result is determined by the matching degree between the topological template of the identifier after affine transformation and the detected identifier position in the current video frame; When the confidence level is lower than the preset confidence threshold, topology matching is performed again or manual intervention is prompted.

7. The method according to claim 1, characterized in that, After determining the target video by using a preset displacement time of a preset location as the end time in the movement trajectory corresponding to each detection marker, the method further includes: Based on the movement trajectory of each detection marker, the movement data of the detection marker is determined, and the movement data includes at least one of displacement, velocity and acceleration; Based on the movement trajectory of each detection marker, a displacement curve of the candidate part is generated; when the region of interest is a human body region, the candidate part includes at least one of the following: head, neck, shoulder, elbow, wrist, spine, hip, knee and ankle. Extract the peak value of a preset type of motion index, wherein the preset type of motion index includes at least one of maximum displacement, peak velocity, and acceleration extreme value; An analysis report is generated based on the movement data, the peak value of the preset type of motion index, and the displacement curve of the candidate part.

8. The method according to claim 1, characterized in that, After determining the first video to be processed by using the moment when the collision signal is obtained as the starting moment, the method further includes: Obtain the in-camera radial distortion coefficients corresponding to the initial collision video; Distortion model correction is performed on the first video to be processed based on the radial distortion coefficients within the camera.

9. An analysis device for collision test videos, characterized in that, include: The first video to be processed determination module uses the moment when the collision signal is obtained as the starting moment to determine the first video to be processed. The region exclusion module excludes non-interested regions from the first video to be processed to obtain the second video to be processed; The detection marker tracking module, in response to all first videos to be processed being converted into second videos to be processed, determines the movement trajectory of each detection marker in each video frame of the second video to be processed and its matching relationship with each part of the target object. The target video determination module determines the target video by using the preset displacement time of a preset part as the end time in the movement trajectory corresponding to each detection marker.

10. An analysis device for collision test videos, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to: perform the analysis method for the crash test video as described in any one of claims 1 to 8.

Images