Mixed flow intersection bicycle riding behavior classification method based on high-dimensional features

Through high-dimensional feature mapping and deep learning technology, combined with self-attention mechanism and K-means clustering, the precise classification of bicycle riding behaviors at mixed-stream intersections is achieved, solving the problem of bicycle abnormal behavior recognition in the existing technology, and improving automatic Driving a vehicle's ability to identify potential risk behaviors reduces the occurrence of traffic accidents.

CN119091353AInactive Publication Date: 2024-12-06GUANGDONG UNIV OF TECH
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411206412.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-30
Publication Date
2024-12-06
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The prior art is difficult to accurately identify the abnormal riding behavior of bicycles in mixed flow intersections, which makes it difficult for autonomous vehicles to predict and prevent traffic accidents. The traditional low-dimensional feature model cannot fully characterize the complex dynamic characteristics of riding behavior.

Method used

Using a hybrid interchange crossing behavior classification method based on high-dimensional features, video data is collected through high-altitude surveillance cameras, bicycle trajectory information is extracted, and high-dimensional features are used using full-connection layer, double-layer long and short-term memory network and self-attention mechanism. Mapping and behavior classification, combining K-means clustering algorithm to divide behavior into two categories: normal and abnormal.

Benefits of technology

It improves the accuracy of bicycle riding behavior classification and abnormal behavior recognition ability, can more accurately capture complex patterns and potential relationships, adapt to complex traffic scenarios, and reduce the occurrence of traffic accidents.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119091353A_ABST
    Figure CN119091353A_ABST
Patent Text Reader

Abstract

The invention relates to a mixed flow intersection bicycle riding behavior classification method based on high-dimensional features. Different from a traditional bicycle riding behavior classification method with low accuracy, the method can efficiently and accurately classify normal and abnormal riding behaviors of bicycles at the mixed flow intersection. Through the method, the automatic driving vehicle can effectively identify potential risk behaviors and take prevention measures in time, so that traffic accidents are reduced. The method comprises the following steps: shooting video data of a mixed flow intersection and extracting track data of a bicycle; mapping the track data of the bicycle to a high-dimensional space through a full connection layer to obtain high-dimensional features; extracting a long-term dependency relationship in the high-dimensional features by using a long-short-term memory network; capturing important moments and features related to the abnormal riding behavior in the long-term dependency relationship of the bicycle through a self-attention mechanism; and finally, classifying the features through a K-means clustering method to obtain a normal riding behavior and an abnormal riding behavior.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of deep learning, traffic big data and abnormal behavior analysis, and in particular to a method for classifying bicycle riding behavior at a mixed-flow intersection based on high-dimensional features. Background Art

[0002] It is very challenging for autonomous vehicles to safely pass through mixed intersections, especially when they encounter bicycles, it is difficult to identify the behavior of bicycles. This is because bicycles need to interact with the complex environment around them, resulting in a large randomness in their behavior, which is considered abnormal behavior. Therefore, it is crucial for autonomous vehicles to accurately identify normal or abnormal riding behavior of bicycles.

[0003] At present, although there are many studies devoted to the modeling of vehicle behavior in mixed-flow intersection scenarios, there are fewer studies on bicycle riding behavior. Most of the existing studies use traditional kinematic models. Traditional low-dimensional features cannot fully characterize the complex dynamic characteristics of riding behavior. The accuracy and generalization ability of the proposed models have certain limitations. Therefore, there is an urgent need for a method that can use high-dimensional features to classify bicycle riding behavior in mixed-flow intersection scenarios, so as to comprehensively solve the problems existing in existing research, so that autonomous driving vehicles can accurately predict the abnormal behavior of surrounding bicycles, effectively identify potential risk behaviors, and take preventive measures in time, thereby reducing the occurrence of traffic accidents. Summary of the invention

[0004] The purpose of the present invention is to provide a method for classifying bicycle riding behavior at mixed-flow intersections based on high-dimensional features, which can accurately and quickly classify bicycle riding behavior at mixed-flow intersections, so that autonomous driving vehicles can accurately predict abnormal behaviors of surrounding bicycles, effectively identify potential risk behaviors, and take preventive measures in time, thereby reducing the occurrence of traffic accidents. In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0005] A method for classifying bicycle riding behavior at mixed-flow intersections based on high-dimensional features comprises the following steps:

[0006] R1: Use an aerial surveillance camera to collect video data of mixed-flow intersections, extract and preprocess bicycle trajectory information. The specific steps are as follows:

[0007] R1-1 Select appropriately located overhead surveillance cameras to ensure that they cover the entire area of ​​the intersection and conduct continuous video capture during peak hours to obtain representative and diverse bicycle data.

[0008] R1-2 Automatically extract the trajectory information of bicycles from the collected video data. Assign a unique ID to each bicycle to maintain its consistency in consecutive frames, thereby ensuring the integrity of the trajectory data. The trajectory data should include the following features of each bicycle in each frame:

[0009] Center position coordinates (x, y): used to determine the specific position of the bicycle in a two-dimensional plane;

[0010] Acceleration (a): represents the rate of change of the bicycle's speed;

[0011] Heading angle (θ): Indicates the angular change in the direction of the bicycle;

[0012] R1-3 smoothes all the extracted bicycle trajectory data to eliminate trajectory discontinuity or jitter caused by detection jitter, ensuring the continuity and naturalness of the trajectory; in addition, it removes abnormal trajectories detected (such as sudden disappearance or unreasonable movement patterns) and processes some lost data through interpolation or completion methods to ensure the continuity and naturalness of the trajectory.

[0013] R2: Implement high-dimensional mapping, feature extraction, and behavior classification of trajectory features. The bicycle trajectory information is mapped to a high-dimensional space through a fully connected layer to obtain high-dimensional space trajectory features, and the long-term dependency features of the bicycle historical trajectory are extracted from the high-dimensional space trajectory features through a long short-term memory network. Then, the self-attention mechanism is used to capture important moments and features related to abnormal bicycle behavior. Finally, these features are classified through the K-means clustering method, and bicycle behavior is divided into two categories: normal riding behavior and abnormal riding behavior. The specific steps are as follows:

[0014] R2-1 Input the bicycle trajectory data preprocessed in step R1-3 into the fully connected layer, map the low-dimensional trajectory information to the high-dimensional space through linear transformation, and capture the complex trajectory patterns and potential associations by introducing higher-dimensional feature representations; after mapping, use nonlinear activation functions after the fully connected layer to introduce nonlinear features to enhance the model's ability to learn complex trajectory relationships, thereby generating high-dimensional space features with stronger representation capabilities.

[0015] R2-2 constructs a two-layer long short-term memory network and inputs the trajectory features mapped to the high-dimensional space in step R2-1 into the two-layer long short-term memory network. The network extracts the long-term pattern of the bicycle trajectory and the continuous change trend of the bicycle acceleration and heading angle by dynamically selecting the information to be retained and forgotten.

[0016] R2-3 uses the output of the double-layer long short-term memory in step R2-2 as the input of the self-attention mechanism. The self-attention mechanism generates attention weights for each time step in the trajectory sequence, highlights key time points and ignores unimportant parts, so that the model can focus on trajectory segments that are highly relevant to abnormal bicycle behavior.

[0017] R2-4 Finally, the K-means clustering algorithm is used to perform cluster analysis on the high-dimensional features extracted in step R2-3. According to the normal and abnormal bicycle riding behavior modes, they can be divided into two clusters. After clustering, each cluster is labeled as a normal or abnormal riding behavior category. The trajectory of an abnormal bicycle is one in which the acceleration change rate and heading angle change rate in the bicycle trajectory data have abnormal fluctuations.

[0018] The above technical solution provides a method for classifying bicycle riding behavior at mixed intersections based on high-dimensional features, which has the following beneficial effects compared with the prior art:

[0019] 1. Improve the classification accuracy of bicycle riding behavior: By introducing high-dimensional feature mapping and deep learning technology, it is possible to more accurately capture the complex patterns and potential associations in bicycle trajectories. Compared with the traditional low-dimensional feature model, this method maps trajectory information to a high-dimensional space through a fully connected layer, combined with a nonlinear activation function, which significantly improves the model's ability to learn complex trajectory relationships, thereby greatly improving the accuracy of behavior classification.

[0020] 2. Enhanced abnormal behavior recognition capability: This method extracts long-term dependent features of trajectories through a two-layer long short-term memory network, and combines it with a self-attention mechanism to effectively capture important moments and features related to abnormal riding behaviors. This mechanism can not only highlight the key points of analysis and filter out irrelevant information, but also accurately identify abnormal fluctuations in acceleration and heading angles in the trajectory, thereby identifying potential dangerous behaviors and improving the recognition rate of abnormal behaviors.

[0021] 3. Adapt to complex traffic scenarios: This method is designed for mixed traffic intersections, taking into account complex traffic environments and diverse behavior patterns. By extracting bicycle trajectory data from aerial surveillance videos and combining deep learning technology to analyze trajectory features, this method can adapt to complex traffic flow scenarios with different traffic densities and different time periods, and has strong robustness and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a logic flow chart of the present invention;

[0023] Figure 2 A mixed flow scene diagram of an intersection in an embodiment of the present invention;

[0024] Figure 3Part of the data after smoothing processing of the present invention;

[0025] Figure 4 The present invention uses KMeans clustering to classify bicycle behavior results;

[0026] Figure 5 This is a bicycle riding trajectory diagram after KMeans clustering in the present invention;

[0027] Figure 6 The cumulative distribution diagram of the bicycle acceleration change in the present invention;

[0028] Figure 7 The cumulative distribution diagram of the change in the bicycle heading angle in the present invention; DETAILED DESCRIPTION

[0029] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0030] This embodiment takes the mixed flow intersection of Haitang Bay in Hefei as an example, and combines the following Figures 1 to 7 , the present invention is further analyzed and explained.

[0031] The flow chart of the classification method of bicycle riding behavior at mixed intersections based on high-dimensional features is as follows: Figure 1 As shown, the following steps are included:

[0032] Step 1: Collect bicycle riding behavior video data in a mixed flow scene near the intersection of Haitang Bay in Hefei City, and pre-process the data. The specific steps are as follows:

[0033] Step 1.1 Select a video of a mixed traffic scene at an intersection during peak traffic hours for data collection, such as Figure 2 Ensure that surveillance cameras cover the entire area of ​​the intersection to obtain representative and diverse bicycle data;

[0034] Step 1.2: Process the collected video data, extract the trajectory of each bicycle, and assign a unique ID to each bicycle to ensure its consistency between consecutive frames. Figure 3 As shown, it contains the center position coordinates (x, y), acceleration (a) and heading angle (θ), which provide the basis for subsequent analysis;

[0035] Step 1.3: Smooth the extracted trajectory data to eliminate trajectory discontinuity caused by camera shaking or detection errors, and ensure the accuracy and naturalness of the data. Figure 3Some of the data after smoothing is shown. In addition, abnormal trajectories (such as trajectories that suddenly disappear or change unreasonably) are detected and removed, and some missing data are interpolated or supplemented. Through these preprocessing steps, the quality and integrity of the trajectory data are ensured.

[0036] Step 2 After completing data collection and preprocessing, further high-dimensional mapping and behavior classification are performed on the bicycle trajectory data. The specific steps are as follows:

[0037] Step 2.1 Input the preprocessed trajectory data into the fully connected layer, and map the low-dimensional trajectory information to a high-dimensional space through linear transformation. By increasing the feature dimension, the present invention can capture more complex trajectory patterns and potential associations. After mapping, a nonlinear activation function is used to enhance the learning ability of the model, thereby generating high-dimensional spatial features with stronger representation capabilities;

[0038] 2.2 Use a two-layer long short-term memory network to process these high-dimensional trajectory features to extract the long-term dependency features of the bicycle's historical trajectory. The long short-term memory network can dynamically capture important patterns in the trajectory, including the continuous change trend of acceleration and heading angle, through the memory and forgetting mechanism. This step enables the model to understand and predict complex cycling behaviors;

[0039] 2.3 Based on the output of the long short-term memory network, the self-attention mechanism is combined to generate attention weights for each time step in the trajectory sequence. The self-attention mechanism can highlight the key time points that are highly related to abnormal behavior and ignore the unimportant parts, so that the model can focus more on potential dangerous behaviors. This process significantly improves the accuracy of the model in identifying abnormal riding behaviors;

[0040] 2.4 The K-means clustering algorithm is used to classify the extracted high-dimensional features and classify bicycle behaviors into normal riding behaviors and abnormal riding behaviors. The classification results are shown in Figure 2. Figure 4 As shown in Figure 1, by analyzing the two clusters, the characteristics of normal and abnormal riding behaviors are clearly identified. Abnormal behavior is usually manifested as abnormal fluctuations in the rate of change of acceleration and heading angle. The riding trajectory after K-means clustering is shown in Figure 1. Figure 5 As shown in the figure, the trajectory characteristics of abnormal behavior can be clearly identified through the trajectory diagram, and it can be distinguished from normal behavior. In order to intuitively display the classification results and the changes in the bicycle riding trajectory, this embodiment visualizes the classified data. Figure 5 The figure shows the bicycle riding trajectory after K-means clustering, and different curves represent normal and abnormal riding behaviors respectively. Figure 6 and Figure 7The cumulative distribution graphs of the bicycle acceleration change and heading angle change are shown respectively. Through these four graphs, the difference between abnormal behavior trajectory and normal behavior trajectory can be clearly observed, especially in terms of acceleration and heading angle changes. Through these graphs, we can deeply understand the behavior pattern of bicycles at intersections, especially in the context of abnormal riding behavior, these distribution graphs reveal potential safety hazards.

[0041] The present invention successfully achieves accurate classification of bicycle riding behavior at mixed-flow intersections by introducing high-dimensional feature mapping, deep learning models and self-attention mechanisms. This method can better understand and predict bicycle riding behavior at mixed-flow intersections, thereby helping autonomous vehicles take more effective measures to prevent potential traffic accidents and improve the safety of the overall traffic environment.

[0042] The embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. For ordinary technicians in this technical field, after knowing the contents recorded in the present invention, they can make several equivalent changes and substitutions without departing from the principle of the present invention. These equivalent changes and substitutions should also be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for classifying bicycle riding behavior at mixed-flow intersections based on high-dimensional features, characterized in that: The following steps are involved: R1: Use an aerial surveillance camera to collect video data of mixed-flow intersections, extract and preprocess bicycle trajectory information. The specific steps are as follows: R1-1 Select appropriately located overhead surveillance cameras to ensure that they cover the entire intersection area and conduct continuous video capture during peak hours to obtain representative bicycle data; R1-2 extracts the trajectory data of bicycles from the collected video data, and ensures the integrity of the trajectory by assigning a unique ID to each bicycle to maintain consistency between consecutive frames; the trajectory data of each bicycle in each frame includes the following features: center position coordinates (x, y), acceleration a, and heading angle θ; R1-3 smoothes all the extracted bicycle trajectory data to eliminate the discontinuity or jitter of the trajectory caused by detection jitter, ensuring the continuity and naturalness of the trajectory; at the same time, it removes abnormal trajectories and interpolates or completes some of the lost data; R2: Map the preprocessed bicycle trajectory information to the high-dimensional space through the fully connected layer to obtain the high-dimensional space trajectory features, and extract the long-term dependency features of the bicycle historical trajectory from the high-dimensional space trajectory features through the long short-term memory network. Then use the self-attention mechanism to capture the important moments and features related to the abnormal bicycle behavior. Finally, use the K-means clustering method to classify these features and divide the bicycle behavior into normal riding behavior and abnormal riding behavior. The specific steps are as follows: R2-1 Input the bicycle trajectory data preprocessed in step R1-3 into the fully connected layer, map the low-dimensional trajectory information to the high-dimensional space through linear transformation, and capture the complex trajectory patterns and potential associations by introducing higher-dimensional feature representations; use nonlinear activation functions after the fully connected layer to introduce nonlinear features to enhance the model's ability to learn complex trajectory relationships; R2-2 construct a two-layer long short-term memory network, and input the trajectory features mapped to the high-dimensional space in step R2-1 into the two-layer long short-term memory network. The network dynamically selects to retain and forget information, thereby extracting the long-term pattern of the bicycle trajectory and the continuous change trend of the bicycle acceleration and heading angle; R2-3 uses the output of the double-layer long short-term memory in step R2-2 as the input of the self-attention mechanism, generates attention weights for each time step in the trajectory sequence through the self-attention mechanism, highlights key time points and ignores unimportant parts, so that the model can focus on trajectory segments that are highly related to abnormal bicycle behavior; R2-4 Finally, the K-means clustering algorithm is used to perform cluster analysis on the high-dimensional features extracted in step R2-3. According to the normal and abnormal bicycle riding behavior modes, they can be divided into two clusters. After clustering, each cluster is labeled as a normal or abnormal riding behavior category. The trajectory of an abnormal bicycle is one in which the acceleration change rate and heading angle change rate in the bicycle trajectory data have abnormal fluctuations.

Citation Information

Patent Citations

  • Intersection abnormal vehicle track recognition and analysis method based on hierarchical clustering

    CN110570658A

  • Multi-modal vehicle trajectory prediction method based on long and short term memory network

    CN113076599A

  • Unsupervised traffic abnormal behavior detection method based on foreground target detection

    CN113221716A

  • Non-motor vehicle driving behavior identification method

    CN116630865A

  • Trajectory prediction method and system based on vehicle behavior recognition

    CN117333847A