Vehicle stop behavior analysis and prediction method based on multidimensional clustering
Through the combination of multi-dimensional clustering and large language model, the problems of poor anti-interference and slow user portrait update in vehicle activity analysis are solved, and efficient vehicle stay behavior analysis and prediction are achieved, ensuring privacy protection and real-time updates are achieved.
Patent Information
- Application Number
- CN202411139064.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-20
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-20
AI Technical Summary
Existing vehicle activity analysis technology faces challenges such as poor anti-interference, slow update of user portraits and privacy protection, making it difficult to effectively analyze and predict vehicle parking behavior.
The multi-dimensional clustering method is adopted to identify the vehicle's activity location through monitoring cameras and machine learning algorithms, and to build user portraits with large language models to achieve full coverage and rapid updates.
It improves the system's anti-interference and identification accuracy, ensures timely updates and privacy protection of user portraits, and provides more accurate and real-time analysis of vehicle parking behavior.
Smart Images

Figure CN118657966B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent transportation and behavior prediction technology, and in particular to a vehicle stop behavior analysis and prediction method based on multidimensional clustering. Background Art
[0002] As an important branch of intelligent transportation technology, vehicle activity location analysis technology analyzes the life and work habits of vehicle owners by counting the places that vehicles frequently visit in different time periods. For example, by identifying the places that vehicles frequently visit, such as home addresses, workplaces, shopping malls, etc., it can reveal the daily activity patterns of vehicle owners. This technology can provide data support and decision-making references for traffic management, urban planning, and commercial marketing.
[0003] Existing technical solutions, such as CN116151489A, disclose a vehicle stop behavior analysis and prediction method and system based on multi-task learning, including the following steps: collecting vehicle GPS and OBD data including vehicle ID after desensitization, trip start time, start longitude, start latitude, end time, end longitude and end latitude; preprocessing the vehicle GPS and OBD data to obtain vehicle stop behavior data including stop location and stop duration; and using a deep recurrent neural network to extract spatiotemporal features of the preprocessed vehicle stop behavior; inputting the spatiotemporal features into a multi-task learning prediction network, and obtaining stop location prediction tasks and stop duration predictions based on the vehicle's historical stop behavior and through the multi-task learning prediction network. The correlation between tasks is used to predict the stop location and the duration of the stop; CN115083161A discloses a method, device, electronic device and readable storage medium for evaluating vehicle stop points, determining multiple vehicles to be evaluated, obtaining the historical vehicle stop points of each vehicle to be evaluated, and obtaining multiple historical vehicle stop points; taking each historical vehicle stop point as the center, querying the set of points of interest within the preset evaluation range, and obtaining the set of points of interest corresponding to each historical vehicle stop point; calculating the repetition rate of the set of points of interest of every two historical vehicle stop points in the multiple historical vehicle stop points, dividing the vehicles to be evaluated corresponding to the historical vehicle stop points whose repetition rate reaches the repetition rate threshold into the same group, and outputting the evaluation result indicating that the vehicles to be evaluated are at the same vehicle stop point. However, when the existing technology is put into practical application, the existing vehicle activity analysis technology generally faces many challenges such as poor anti-interference, slow user portrait update, and privacy protection. Therefore, a new vehicle stop behavior analysis and prediction method is urgently needed. Summary of the invention
[0004] In order to solve the above-mentioned problems such as poor anti-interference, slow update of user portraits, and privacy protection, the present invention proposes a vehicle stop behavior analysis and prediction method based on multidimensional clustering. It uses surveillance cameras to identify the frequented places of car owners in different time periods through machine learning algorithms such as target detection, optical character recognition, unsupervised clustering and scene recognition, and uses a large language model to create a user portrait for each car owner.
[0005] In order to achieve the above object, the present invention adopts the following technical solution:
[0006] A vehicle stop behavior analysis and prediction method based on multidimensional clustering, the method specifically includes the following processes:
[0007] S1. Image acquisition: Multi-camera linkage is adopted to link the surveillance cameras at the parking lot entrance and the adjacent areas through edge devices to record the activity locations of vehicles, achieve full coverage in the process of license plate recognition and activity location analysis, reduce occlusion and blind spots, and divide the video frames into four time types according to the shooting time: daytime, nighttime, weekdays, and holidays according to the camera's timestamp;
[0008] S2, license plate recognition: using the automatic license plate recognition system to identify the license plate number in the video frame;
[0009] S3. Multidimensional clustering: Use K-mean multidimensional clustering algorithm to perform unsupervised clustering on video frames in different time periods, and output a cluster label table to record which cluster each video frame belongs to and the size of each cluster. Select the largest R clusters to facilitate subsequent analysis of the places that car owners often visit during different time periods during the day / night, weekdays / holidays;
[0010] S4, scene recognition: Based on the principles of typicality and diversity, an equal number of video frames are extracted from the R clusters obtained in step S3 as candidate frames, and the AlexNet model pre-trained on the large scene dataset Places365 is used to perform scene recognition on the candidate frames, and the scene categories of the R clusters are predicted according to the principle of minority obeys majority;
[0011] S5. User portrait: Based on the license plate number, time type and scene type, the Chinese large model chatGLM2 is used to create an accurate personalized portrait for each car owner.
[0012] As a further technical solution of the present invention, the format of the time type in step S1 is YYYY-MM-DD HH:MM:SS, i.e. year, month, day, hour, minute, and second. 6 a.m. to 6 p.m. is defined as daytime, 6 p.m. to 6 a.m. the next day is defined as night, and working days and holidays are distinguished according to statutory holidays.
[0013] As a further technical solution of the present invention, the specific process of step S2 is:
[0014] S21. Vehicle detection: input video frame , the target detection model YOLOv2 is used to detect vehicles in noisy scenes, and the vehicle bounding box is used represents the position of the vehicle in the image, where Represent the center position and length and width of the vehicle bounding box respectively;
[0015] S22, license plate detection and correction: Use the distorted planar object detection network WPOD-Net to find the area where the license plate is located in the vehicle bounding box by traversing. ,in are the coordinates of the four vertices of the license plate bounding box expressed in the clockwise direction. Formula (1) is used to adjust the area where the license plate is located into a rectangle in the front view state through affine transformation to correct the license plate angle and obtain the corrected area. Specifically, for a license plate detection frame with a height of H and a width of W, the dimension of the feature map output by WPOD-Net is M×N×8, where M = H / N S and N = W / N S , N S =4 represents the step size of the network. For each point cell (m, n) in the feature map, there are 8 values, and the last six values v3 to v8 are used to implement the affine transformation:
[0016] (1),
[0017] Among them, max(v3, 0) and max(v6, 0) make v3 and v6 non-negative, thus ensuring that the diagonal is positive, that is, avoiding the mirror image or excessive rotation of the license plate, and the constant , , , , ;
[0018] S23, License Plate Recognition: Use the character detection and recognition model OCR-Net to extract the license plate number in the correction area in the form of text.
[0019] As a further technical solution of the present invention, the specific process of step S3 is:
[0020] Use N= , , ] represents the number of video frames in four time types: daytime, nighttime, weekdays and holidays. Input the deep neural network AlexNe to get the feature vector f output by the feature extractor. Each image in the video frame There is a corresponding eigenvector ,in , E represents the dimension of the feature vector. The feature vector is subjected to K-means cluster analysis. The cluster analysis includes the following six steps:
[0021] S31, Initialization: Randomly select K feature vectors as initial cluster centers, and use Indicates that A represents the maximum number of iterations;
[0022] S32, allocation: each feature vector is assigned to its nearest cluster center. With cluster center The Euclidean distance between ,in, Represents the feature vector The eigenvalue of the mth dimension;
[0023] S33, Update: Calculate the new position of each cluster center, that is, the mean of all video frame feature vectors in the cluster, using Indicates that is assigned to the cluster center The feature set cluster, It represents the number of eigenvectors in the cluster, that is, the size of the cluster;
[0024] S34, iteration: repeating step S32 and step S33 until the cluster center no longer changes significantly, or a preset number of iterations A is reached;
[0025] S35. Cluster label table: Generate cluster label table , record the cluster to which each feature belongs, where is the video frame The index of the cluster to which it belongs;
[0026] S36, cluster size: Indicates the size of each cluster.
[0027] As a further technical solution of the present invention, the specific process of step S4 is:
[0028] S41, sorting the clusters according to the number of features, and selecting R clusters with the largest number of features;
[0029] S42. Use the AlexNet model to perform scene recognition on the candidate set, where the candidate set consists of two parts: random frames and central frames. The random frames randomly extract B features from R clusters to form a feature set of size R×B. The central frame selects the feature closest to the cluster center. Each cluster center selects B features to form a feature set of size R×B. The size of the candidate set is R×2B. Each feature in the candidate set is directly input into the classifier of the AlexNet model to predict the scene category of the video frame, and according to the principle of minority obeys majority, it is determined which type of scene each cluster represents.
[0030] As a further technical solution of the present invention, the process of obtaining the user image in step S5 is expressed as U=V+Model(T,Z), where V is the license plate number, Model represents the Chinese large model chatGLM2; time type Representing daytime, nighttime, weekdays and holidays respectively; Represent the scene types respectively; Represents user tags respectively.
[0031] Compared with the prior art, the present invention has the following beneficial effects:
[0032] 1. Anti-interference: Through multi-camera linkage and edge device processing, all-round monitoring of the parking lot entrance and adjacent areas is achieved, effectively reducing occlusion and blind spots, which is conducive to license plate detection and surrounding scene perception. At the same time, the automatic license plate recognition system can accurately identify the license plate even in tilted or noisy background conditions, significantly improving the system's anti-interference and recognition accuracy.
[0033] 2. Portrait update: It uses unsupervised clustering algorithms, pre-trained scene classification models and large language models. There is no need to retrain the model, which effectively saves computing power and time. It can update user portraits in a timely manner and ensure that the updated portraits accurately reflect the latest status of the car owner.
[0034] 3. Privacy protection: When performing unsupervised clustering of vehicle activity locations in different time periods and scene classification of locations, no labeled vehicle data is required, which is conducive to protecting the privacy of vehicle owners and preventing the leakage of their personal information. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present disclosure and do not constitute improper limitations on the present invention.
[0036] Figure 1 A schematic flow chart of the vehicle stop behavior analysis and prediction method based on multidimensional clustering provided by the present invention. DETAILED DESCRIPTION
[0037] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0038] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.
[0039] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, device (device), product or equipment that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment.
[0040] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.
[0041] Example:
[0042] like Figure 1 As shown, this embodiment provides a vehicle stop behavior analysis and prediction method based on multidimensional clustering, comprising the following steps:
[0043] S1. Image collection:
[0044] This embodiment achieves 360-degree comprehensive monitoring of the parking lot entrance and its adjacent areas through camera layout to ensure no blind spots or obstructions. The selected cameras have high resolution and excellent low-light performance, and can adapt to lighting conditions in different time periods such as daytime, nighttime, and holidays, clearly capture vehicle and scene features, and the dynamic capture capability of the camera ensures real-time tracking of vehicle movement. The camera at the entrance of the parking lot focuses on license plate recognition, while the cameras in the surrounding area record scene information of vehicles passing by. These image data are transmitted to the edge device in real time to parse the camera's timestamp and classify and store the video frames according to daytime, nighttime, weekdays, and holidays. The time format is (YYYY-MM-DD HH:MM:SS), that is, year-month-day-hour-minute-second. Six o'clock in the morning to six o'clock in the evening is defined as daytime, and statutory holidays are used to distinguish between weekdays and holidays.
[0045] S2. License Plate Recognition:
[0046] This embodiment uses an automatic license plate recognition system to recognize the license plate number in the video frame; the specific process is:
[0047] S21. Vehicle detection: input video frame , the target detection model YOLOv2 is used to detect vehicles in noisy scenes, and the vehicle bounding box is used represents the position of the vehicle in the image, where Represent the center position and length and width of the vehicle bounding box respectively;
[0048] S22, license plate detection and correction: Use the distorted planar object detection network WPOD-Net to find the area where the license plate is located in the vehicle bounding box by traversing. ,in are the coordinates of the four vertices of the license plate bounding box expressed in the clockwise direction. Formula (1) is used to adjust the area where the license plate is located into a rectangle in the front view state through affine transformation to correct the license plate angle and obtain the corrected area. Specifically, for a license plate detection frame with a height of H and a width of W, the dimension of the feature map output by WPOD-Net is M×N×8, where M = H / N S and N = W / N S , N S =4 represents the step size of the network. For each point cell (m, n) in the feature map, there are 8 values, and the last six values v3 to v8 are used to implement the affine transformation:
[0049] (1),
[0050] Among them, max(v3, 0) and max(v6, 0) make v3 and v6 non-negative, thus ensuring that the diagonal is positive, that is, avoiding the mirror image or excessive rotation of the license plate, and the constant , , , , ;
[0051] S23, License Plate Recognition: Use the character detection and recognition model OCR-Net to extract the license plate number in the correction area in the form of text.
[0052] S3, multidimensional clustering:
[0053] This embodiment uses the K-mean multidimensional clustering algorithm to perform unsupervised clustering on video frames in different time periods, and outputs a cluster label table to record which cluster each video frame belongs to and the size of each cluster. The largest R clusters are selected to facilitate subsequent analysis of the places frequently visited by car owners during different time periods during the day / night, weekdays / holidays. Specifically, N= , , ] represents the number of video frames in four time types: daytime, nighttime, weekdays and holidays. Input the deep neural network AlexNe, which has been pre-trained on the large scene dataset Places365 and does not need to be retrained. Then get the feature vector f output by the feature extractor. Each image in the video frame There is a corresponding eigenvector ,in , E represents the dimension of the feature vector. The feature vector is subjected to K-means cluster analysis. The cluster analysis includes the following six steps:
[0054] S31, Initialization: Randomly select K feature vectors as initial cluster centers, and use Indicates that A represents the maximum number of iterations;
[0055] S32, allocation: each feature vector is assigned to its nearest cluster center. With cluster center The Euclidean distance between ,in, Represents the feature vector The eigenvalue of the mth dimension;
[0056] S33, Update: Calculate the new position of each cluster center, that is, the mean of all video frame feature vectors in the cluster, using Indicates that is assigned to the cluster center The feature set cluster, It represents the number of eigenvectors in the cluster, that is, the size of the cluster;
[0057] S34, iteration: repeating step S32 and step S33 until the cluster center no longer changes significantly, or a preset number of iterations A is reached;
[0058] S35. Cluster label table: Generate cluster label table , record the cluster to which each feature belongs, where is the video frame The index of the cluster to which it belongs;
[0059] S36, cluster size: Indicates the size of each cluster.
[0060] S4, scene recognition:
[0061] Based on the principles of typicality and diversity, an equal number of video frames are extracted from the R clusters obtained in step S3 as candidate frames. The AlexNet model pre-trained on the large scene dataset Places365 is used to perform scene recognition on the candidate frames, and the scene categories of the R clusters are predicted according to the principle of minority obeys majority, specifically:
[0062] First, the clusters are sorted by the number of features, and then the R clusters with the largest number of features are selected. In order to balance typicality and diversity, this embodiment uses the AlexNet model to perform scene recognition on the candidate set, where the candidate set consists of two parts: random frames and center frames. The random frames randomly extract B features from R clusters to form a feature set of size R×B, which ensures the diversity of features; the center frame selects the feature closest to the cluster center, and each cluster center selects B features to form a feature set of size R×B, which ensures the typicality of the features. Therefore, the size of the candidate set is R×2B, and each feature in the candidate set is directly input into the classifier of the AlexNet model to predict the scene category of the video frame. The model has been pre-trained on the large scene dataset Places365. There are 2E features in the candidate set in each cluster. According to the principle of minority obeys majority, it is judged which type of scene each cluster represents, such as shopping malls, gymnasiums, schools, etc.
[0063] S5. User portrait:
[0064] According to the license plate number, time type and scene type, the Chinese big model chatGLM2 is used to create an accurate personalized portrait for each car owner. The reason for choosing chatGLM2 is that it not only has good performance, but is also one of the few Chinese big models that can be deployed locally and provide open source APIs. Specifically, through the chatGLM2 API, the corresponding user label U can be quickly generated based on the license plate number V and its activity records at a specific time type T and location type Z, thereby building a personalized portrait of the car owner. From the perspective of privacy protection, the license plate number will not be input into the big model, but will be connected to the output user label. Therefore, the complete process of user portrait is represented by U=V+Model(T,Z), where Model represents the Chinese big model chatGLM2; time type Representing daytime, nighttime, weekdays and holidays respectively; They represent residential areas, shopping malls and other scene types respectively; They represent user labels such as "overtime maniac" or "natural scenery lover". When the time type T and scene type Z are input into the large model, this embodiment uses the prompt template "The car owner likes to stay in scene [Z] within time period [T]. Please describe his user portrait in five words."
[0065] The vehicle stop behavior analysis and prediction method based on multidimensional clustering proposed in this embodiment adopts unsupervised clustering, which will not involve the privacy of the car owner, and can also mine and utilize the data patterns and structures unique to each activity location during the clustering process. Moreover, compared with directly using the classification algorithm, first using the clustering algorithm to cluster the scene is conducive to reducing the impact of noise and outliers, that is, unsupervised clustering of activity locations in different time periods is performed to find the most frequently occurring locations, and then the image classification model is used to identify the scene type of the location. Finally, based on the locations frequently visited by the vehicle and the corresponding time, a more accurate and faster updating user portrait is provided for the car owner, providing personalized services that are more popular with the car owner. At the same time, this embodiment helps to improve the real-time warning of traffic congestion, or be used for more economical and comfortable urban planning.
[0066] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
[0067] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.
Claims
1. A vehicle stop behavior analysis and prediction method based on multidimensional clustering, characterized in that: The specific process includes the following: S1. Image acquisition: Using multi-camera linkage, the edge devices are used to link the surveillance cameras at the parking lot entrance and the adjacent area. Based on the camera timestamps, the video frames are divided into four time types: daytime, nighttime, weekdays, and holidays. S2, license plate recognition: using the automatic license plate recognition system to identify the license plate number in the video frame; S3, multidimensional clustering: Use K-mean multidimensional clustering algorithm to perform unsupervised clustering on video frames in different time periods, and output a cluster label table to record which cluster each video frame belongs to, and the size of each cluster, and select the largest R clusters; S4, scene recognition: Based on the principles of typicality and diversity, an equal number of video frames are extracted from the R clusters obtained in step S3 as candidate frames, and the AlexNet model pre-trained on the large scene dataset Places365 is used to perform scene recognition on the candidate frames, and the scene categories of the R clusters are predicted according to the principle of minority obeys majority; S5. User portrait: Based on the license plate number, time type and scene type, the Chinese large model chatGLM2 is used to create an accurate personalized portrait for each car owner.
2. The vehicle stop behavior analysis and prediction method based on multidimensional clustering according to claim 1 is characterized in that: The time type in step S1 is in the format of YYYY-MM-DD HH:MM:SS, i.e. year, month, day, hour, minute, and second. 6:00 a.m. to 6:00 p.m. is defined as daytime, 6:00 p.m. to 6:00 a.m. the next day is defined as nighttime, and working days and holidays are distinguished according to statutory holidays.
3. The vehicle stop behavior analysis and prediction method based on multidimensional clustering according to claim 2 is characterized in that: The specific process of step S2 is: S21. Vehicle detection: input video frame , the target detection model YOLOv2 is used to detect vehicles in noisy scenes, and the vehicle bounding box is used represents the position of the vehicle in the image, where Represent the center position and length and width of the vehicle bounding box respectively; S22, license plate detection and correction: Use the distorted planar object detection network WPOD-Net to find the area where the license plate is located in the vehicle bounding box by traversing. ,in are the coordinates of the four vertices of the license plate bounding box expressed in the clockwise direction. Formula (1) is used to adjust the area where the license plate is located into a rectangle in the front view state through affine transformation to correct the license plate angle and obtain the corrected area. Specifically, for a license plate detection frame with a height of H and a width of W, the dimension of the feature map output by WPOD-Net is M×N×8, where M = H / N S and N = W / N S , N S =4 represents the step size of the network. For each point cell (m, n) in the feature map, there are 8 values, and the last six values v3 to v8 are used to implement the affine transformation: (1), Among them, max(v3, 0) and max(v6, 0) make v3 and v6 non-negative, thus ensuring that the diagonal is positive, that is, avoiding the mirror image or excessive rotation of the license plate, and the constant , , , , ; S23, License Plate Recognition: Use the character detection and recognition model OCR-Net to extract the license plate number in the correction area in the form of text.
4. The vehicle stop behavior analysis and prediction method based on multidimensional clustering according to claim 3 is characterized in that: The specific process of step S3 is: Use N= , , ] represents the number of video frames in four time types: daytime, nighttime, weekdays and holidays. Input the deep neural network AlexNe to get the feature vector f output by the feature extractor. Each image in the video frame There is a corresponding eigenvector ,in , E represents the dimension of the feature vector. The feature vector is subjected to K-means cluster analysis. The cluster analysis includes the following six steps: S31, Initialization: Randomly select K feature vectors as initial cluster centers, and use Indicates that A represents the maximum number of iterations; S32, allocation: each feature vector is assigned to its nearest cluster center. With cluster center The Euclidean distance between ,in, Represents the feature vector The eigenvalue of the mth dimension; S33, Update: Calculate the new position of each cluster center, that is, the mean of all video frame feature vectors in the cluster, using Indicates that is assigned to the cluster center The feature set cluster, It represents the number of eigenvectors in the cluster, that is, the size of the cluster; S34, iteration: repeating step S32 and step S33 until the cluster center no longer changes significantly, or a preset number of iterations A is reached; S35. Cluster label table: Generate cluster label table , record the cluster to which each feature belongs, where is the video frame The index of the cluster to which it belongs; S36, cluster size: Indicates the size of each cluster.
5. The vehicle stop behavior analysis and prediction method based on multidimensional clustering according to claim 4 is characterized in that: The specific process of step S4 is: S41, sorting the clusters according to the number of features, and selecting R clusters with the largest number of features; S42. Use the AlexNet model to perform scene recognition on the candidate set, where the candidate set consists of two parts: random frames and central frames. The random frames randomly extract B features from R clusters to form a feature set of size R×B. The central frame selects the feature closest to the cluster center. Each cluster center selects B features to form a feature set of size R×B. The size of the candidate set is R×2B. Each feature in the candidate set is directly input into the classifier of the AlexNet model to predict the scene category of the video frame, and according to the principle of minority obeys majority, it is determined which type of scene each cluster represents.
6. The vehicle stop behavior analysis and prediction method based on multidimensional clustering according to claim 5 is characterized in that: The process of obtaining the user image in step S5 is represented as U=V+Model(T,Z), where V is the license plate number, Model represents the Chinese large model chatGLM2; time type Representing daytime, nighttime, weekdays and holidays respectively; Represent the scene types respectively; , , } respectively represent user tags.
Citation Information
Patent Citations
Vehicle parking point evaluation method and device, electronic equipment and readable storage medium
CN115083161A
Vehicle parking behavior analysis and prediction method and system based on multi-task learning
CN116151489A
Automatic generating method for multi-lane vehicle track space-time diagram
CN103903019A
Behavior recognition method and device based on SURF interest points
CN103955682A
Vehicle behavior analysis method and system
CN105632175A
