Vehicle trajectory anomaly detection method based on multi-feature fusion

Abstract

The invention discloses a vehicle trajectory anomaly detection method based on multi-feature fusion, and the method comprises the steps: obtaining vehicle trajectory data, and carrying out the preprocessing of the vehicle trajectory data; performing trajectory feature extraction on the preprocessed vehicle trajectory data set, determining a normal interval of each index in the data set by adopting an improved sectional 3 sigma principle, and completing initial screening of abnormal data; according to the method, vehicle track characteristics based on a space-time sequence and statistics are fused, multi-angle track density characteristics are introduced, space-time characteristics, statistical characteristics and the multi-angle track density characteristics are subjected to dynamic weight to form new characteristics, and the new characteristics are fused through adaptive weight to become input of a residual network; and optimizing the residual network model by using a reward and punishment mechanism considering a frequency factor of continuous abnormal points to realize anomaly detection. According to the method, various features and abnormal conditions of the vehicle track are comprehensively considered, the function of feature fusion is fully played, and the abnormal vehicle track can be identified more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle trajectories, and in particular to a vehicle trajectory anomaly detection method based on multi-feature fusion. Background Art

[0002] In modern logistics and transportation, vehicle trajectory anomaly detection is crucial. Accurately identifying abnormal vehicle trajectories can effectively improve logistics and transportation safety, optimize logistics and transportation management, and improve logistics and transportation efficiency.

[0003] Although the existing vehicle trajectory anomaly detection methods have tried to use multiple features to a certain extent, there are still many shortcomings. On the one hand, some methods are not deep and comprehensive enough when fusing multiple features. Although the basic features such as vehicle speed and acceleration are taken into account, the comprehensive application of spatiotemporal features, statistical features and multi-angle trajectory density features is relatively limited. For example, when fusing different features, it may simply be a linear combination without fully exploring the inherent correlation and complementarity between the features. There is also a lack of dynamic adjustment mechanism for the changes in the importance of each feature in different traffic scenarios, resulting in low detection accuracy in complex and changeable actual traffic environments.

[0004] On the other hand, insufficient attention has been paid to multi-angle trajectory density features. Behavioral frequency features, temporal density features, and spatial density features can reflect the behavior patterns of vehicles in a specific area from different angles, but many current methods fail to fully utilize these features. Behavioral frequency features can help detect drivers' bad driving habits and potential safety hazards, temporal density features can help discover potential areas and time periods for traffic congestion, and spatial density features are of great value in analyzing traffic flow and planning road resources. However, the application of these features in existing methods is often not in-depth and systematic enough.

[0005] In summary, there is a need for a more complete vehicle trajectory anomaly detection method based on multi-feature fusion, which can give full play to the advantages of each feature, deeply explore the intrinsic connection between the features, and attach importance to the use of multi-angle trajectory density features to improve the accuracy, reliability and adaptability of vehicle trajectory anomaly detection. Summary of the invention

[0006] The present invention provides a vehicle trajectory anomaly detection method based on multi-feature fusion, in order to solve the problem that most of the current vehicle trajectory anomaly detection methods have insufficient feature fusion and lack of attention to multi-angle trajectory features.

[0007] To achieve the above object, the present invention provides the following technical solution: a vehicle trajectory anomaly detection method based on multi-feature fusion, comprising the following steps:

[0008] S1. Send data to a designated server through the built-in device on the vehicle, then obtain data from the server to obtain vehicle trajectory data;

[0009] S2, preprocessing the vehicle trajectory data, including data cleaning, data smoothing, noise removal and time alignment;

[0010] S3, extract trajectory features from the pre-processed vehicle trajectory data set, use the improved segmented 3σ principle to determine the threshold standards of each indicator in the vehicle trajectory data set, mark the vehicle trajectory data outside the threshold as abnormal trajectory data, complete the preliminary screening and elimination of abnormal trajectory data and the marking of abnormal behavior types, and perform further abnormal trajectory detection on the remaining data set;

[0011] S4, weighted fusion of spatiotemporal features, statistical features and multi-angle trajectory density features into new features for the detection model to learn, so as to detect abnormal vehicle trajectories;

[0012] S5, training and detection module, inputs historical abnormal vehicle trajectory data and fusion features into the residual network model to perform abnormal trajectory detection on the vehicle trajectory.

[0013] Preferably, in step S1, the vehicle-mounted sensor is used to interact and collect the road condition information of the road section where the vehicle is located in real time, generate detailed lane information and the range and area of ​​the geographical area and the number of vehicles in the area, and form relative coordinates according to the current coordinate system and time point and save them together with the road condition information to form a new trajectory data coordinate point (x i ,y i , t i , J i );

[0014] Among them, the vehicle trajectory data includes global geographic coordinates, vehicle speed and acceleration;

[0015] Among them, x and y are global relative geographic coordinates, t is the current time point of the trajectory, J represents the road condition of the current trajectory point, and i represents the current trajectory point on the vehicle trajectory.

[0016] Preferably, in step S2: data cleaning, a cleaning method is adopted to perform cleaning according to the distribution characteristics of the data, so as to remove duplicate information, blank information and erroneous information in the data set and maintain the uniformity of the characteristics of the data set.

[0017] Preferably, in the step S3, the improved segmented 3σ principle divides the vehicle's driving trajectory according to different road sections, selects vehicle coordinates, collection time and lane information as characteristic indicators, observes the group characteristics of each indicator of the time and space conditions included in each data set, and the group characteristics include spatiotemporal characteristics, statistical characteristics, and multi-angle trajectory density characteristics; and obtains the data intervals where most of the indicators are located for each road section, and the data that does not appear in these intervals shows characteristics different from other data. When a screening indicator in the data record is within the abnormal interval, the data record is marked as abnormal behavior data.

[0018] Preferably, the multi-angle trajectory density features include behavior frequency features, time density features and space density features;

[0019] Among them, the behavior frequency f behievr.ρ.i It indicates the number of specific abnormal behaviors that occur when traversing to the i trajectory point in a specific ρ area;

[0020] Among them, the time density

[0021] Among them, f time.k.i represents the time density of the i-th trajectory point in the k-th specific time period, T k represents the time point set of vehicles in the area during the kth specific time period, T total，k represents the total length of the kth specific time period;

[0022] Among them, the spatial density

[0023] Among them, f space.ρ.i represents the spatial density of the ith trajectory point in the ρth specific geographic area, N vehicle，ρ represents the number of vehicles in the ρth specific geographical area, A region，ρ represents the area of ​​the ρth specific geographical region.

[0024] Preferably, in step S4, the new feature is a comprehensive feature obtained by fusing multiple features.

[0025]

[0026] Among them, v ρi is the speed characteristic, a ρi is the acceleration characteristic and θ ρi The directional feature is a sub-feature of the spatiotemporal feature. is the average speed characteristic, D ρi is the driving distance characteristic and T ρstop，piThe residence time feature is a sub-feature of the statistical feature; the weights of the internal sub-features in the spatiotemporal feature, statistical feature and spatial density feature are evenly distributed; α, β, and γ are the weight coefficients of the spatiotemporal feature, statistical feature and multi-angle trajectory density feature, respectively, and α+β+γ=1 is satisfied. Users can score each scene factor according to their needs in different scenarios and make dynamic adjustments according to their needs.

[0027] Preferably, α, β, and γ are weight coefficients, and users can score each scenario factor in different scenarios according to their needs, and make dynamic adjustments including:

[0028] S41. Construct feature matrix (H)

[0029]

[0030] Where i is the number of trajectory points, std i represents the standard deviation of each feature corresponding to the i-th trajectory point, so that some highly sensitive features are not underestimated; impi i The importance ranking value of each feature corresponding to the i-th trajectory point is used to determine the importance of each feature, and a larger weight coefficient is assigned to a feature with higher importance, and a smaller weight coefficient is assigned to a feature with lower importance;

[0031] S42. Build scene matrix (G)

[0032]

[0033] Where C represents the scene factor, a represents the number of scene factors, and b represents the number of scenes. In addition, a scene factor C representing the sensitivity to the standard deviation is added. std and scenario factor C, which indicates the importance of ranking imp ;

[0034] S43. Calculate the weight matrix (W)

[0035] First, define the function of standard deviation and importance ranking

[0036] f(std, imp)=std*imp+std+imp

[0037] W=H*(G T *diago(f(std1, imp1), f(std2, imp2),, f(std i , imp i )))

[0038] where diago(f(std1, imp1), f(std2, imp2), f(std i , impi )) means f(std i , imp i ) is a diagonal matrix with diagonal elements; in different scenarios, score each scenario factor according to needs, and obtain dynamic α, β, and γ weight coefficients respectively.

[0039] Preferably, in the step S5, the trajectory detection of the vehicle trajectory includes four types of anomalies, including detour anomalies, speed anomalies, lane change anomalies and stop anomalies, and a reward mechanism is defined. The input features are processed by the residual network to obtain preliminary trajectory anomaly detection results; the reward mechanism adjusts the network parameters according to the comparison between the detection results and the actual results, optimizes the detection performance, uses the reward function to evaluate the accuracy of the detection results, and obtains the parameter combination with the highest score of the residual network model. The trajectory data to be detected is detected by the optimized residual network model, and the anomaly type is output to complete the abnormal detection of the vehicle trajectory.

[0040] Preferably, in step S5,

[0041] S51. Input the fusion features and historical trajectory data with abnormal type labels into the residual network, design a reward and punishment mechanism, and the reward and punishment score R = R - *n - -R + *n + ;

[0042] Where R - Represents the penalty value, which is the penalty value given when the residual network mistakenly detects the vehicle trajectory as normal but is actually abnormal, or mistakenly detects the vehicle trajectory as abnormal but is actually normal or an undetected abnormal type; R + Represents the reward value, which is the reward value given when the residual network accurately detects that the vehicle trajectory is normal and is actually normal, or accurately detects that the vehicle trajectory is a certain abnormal type and is actually the abnormal type; n - and n + Represent the number of times of being punished and rewarded respectively;

[0043] S52. Define the loss function: loss function L

[0044]

[0045] Among them, q represents the actual vehicle trajectory label, Represents the detection result of the residual network, μ is the balance coefficient, n abnormal The number of consecutive abnormal points, τ is the adjustment coefficient;

[0046] Repeat the above steps and continuously optimize the parameters of the residual network until the comprehensive performance of the output results reaches a stable level to improve the prediction accuracy of the model. After the residual network is trained, input the trajectory data to be detected, output the abnormal trajectory type, and complete the abnormal detection of the vehicle trajectory.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The present invention comprehensively considers various characteristics and abnormal situations of vehicle trajectories, gives full play to the role of feature fusion, and can more accurately identify abnormal vehicle trajectories.

[0049] 2. The present invention can give full play to the advantages of each feature, deeply explore the internal connection between the features, and attach importance to the use of multi-angle trajectory density features to improve the accuracy, reliability and adaptability of vehicle trajectory anomaly detection.

[0050] 3. In the present invention, spatiotemporal features, statistical features and trajectory density features are weighted and fused into new features, which can improve the sensitivity of anomaly detection, adapt to different data distributions, enhance flexibility and customizability, and at the same time highlight key features, reduce noise interference and adapt to different task requirements, thereby optimizing feature weight calculation to improve the accuracy and pertinence of vehicle trajectory anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0052] In the attached picture:

[0053] Figure 1 It is a structural schematic diagram of a vehicle trajectory anomaly detection method based on feature fusion of the present invention;

[0054] Figure 2 It is a flow chart of the trajectory initial screening of the improved segmented 3σ principle in the present invention;

[0055] Figure 3 It is a flow chart of feature weight calculation in the present invention. DETAILED DESCRIPTION

[0056] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0057] Example: Figure 1 As shown, a vehicle trajectory anomaly detection method based on multi-feature fusion includes the following steps:

[0058] S1. The logistics company system sends data to the designated server through the built-in devices on the vehicle, such as mobile network modules and GPS devices, and then obtains data from the server, while generating detailed lane information and the scope of the geographical area, the number of vehicles in the area, and the area, etc., and saves the relative coordinates according to the current coordinate system and time point, and makes the time exist in the form of a timestamp to facilitate subsequent operations;

[0059] In step S1, it is necessary to use the on-board sensor to interact with the traffic management system to collect the road condition information of the road section where the vehicle is located in real time. The vehicle trajectory data includes global geographic coordinates, vehicle speed, acceleration; generate detailed lane information, the scope and area of ​​the geographic area, and the number of vehicles active in the area, and form relative coordinates according to the current coordinate system and time point, and save them at a frequency of 5 frames per second. The road condition information is saved together with the vehicle trajectory data to form a new trajectory data coordinate point (x i ,y i , t i , J i ), x and y are global relative geographic coordinates, t is the current time point of the trajectory, J represents the road condition of the current trajectory point, and i represents a trajectory point on the vehicle trajectory.

[0060] S2, preprocessing the vehicle trajectory data, including data cleaning, data smoothing, noise removal and time alignment, to obtain a processed trajectory data set;

[0061] In step S2, the data cleaning method used is to clean the data according to the distribution characteristics of the data, which is used to remove duplicate information, blank information and erroneous information in the data set, mainly including removing missing values, removing data showing errors, and removing useless fields, so that the data set shows unified characteristics and the data set is filtered into high-quality data; data smoothing, using the SG filter method to smooth the vehicle trajectory data and remove the original noise in the data set; time alignment, ensuring the uniformity of timestamps of different data sources, facilitating subsequent data analysis and processing.

[0062] S3, extract the characteristics of the vehicle trajectory data set, and then determine the threshold standards of each indicator in the vehicle trajectory data set according to the improved segmented "3σ principle", mark the vehicle trajectory data outside the threshold as abnormal trajectory data, complete the preliminary elimination and screening of abnormal trajectory data and the marking of abnormal behavior types, and use the remaining data set after the screening as a new vehicle trajectory data set;

[0063] In step S3, if Figure 2As shown in the figure, the improved segmented "3σ principle" refers to dividing the vehicle's driving trajectory into different sections such as expressway sections, urban congested sections, and suburban sections, selecting vehicle coordinates, collection time, lane information, etc. as feature indicators, observing the group characteristics of each indicator of the time and space conditions included in each data set, including spatiotemporal characteristics, statistical characteristics, and multi-angle trajectory density characteristics, and obtaining the data intervals where most of the indicators are located for each section. Data that does not appear in these intervals show characteristics different from other data. When a screening indicator in the data record is within the abnormal interval, the data record is marked as abnormal behavior data.

[0064] Among them, the sub-features with normal regional ranges involved in the spatiotemporal features and statistical features are:

[0065] The space-time sub-characteristics include velocity characteristics (v ρi ), acceleration characteristics (a ρi ) and directional characteristics (θ ρi ).

[0066] The speed

[0067] where v ρi represents the speed of the ith trajectory point on the ρ section, x i ,y i are the global geographic coordinates of the i-th point, t i represents the data collection time of the i-th point, and ρ represents the road section number.

[0068] According to the improved segmented “3σ principle”, the normal range is:

[0069]

[0070] n is the total number of trajectories traversed to point i. ρi If it exceeds the normal range, it may be judged as abnormal speed and initially eliminated.

[0071] The acceleration The mean value of acceleration Mean standard deviation of acceleration

[0072] According to the improved segmented “3σ principle”, the normal range is:

[0073] [a μρi -3a σρi , a μρi +3a σρi ]

[0074] If a ρiIf it exceeds the normal range, it may be judged as abnormal speed or abnormal lane change and initially eliminated.

[0075] The direction

[0076] According to the improved segmented “3σ principle”, the normal range is:

[0077] [θ μρi -3θ σρi ,θ μρi +3θ σρi ]

[0078] If θ ρi If it is beyond the normal area, it means that the vehicle changes lanes frequently, and it may be judged as abnormal lane change and preliminarily eliminated.

[0079] The statistical sub-features include average speed features Driving distance characteristics (D ρi ) and residence time characteristics (T ρstop，pi ),

[0080] The average speed

[0081] According to the improved segmented “3σ principle”, the normal range is:

[0082]

[0083] like If it is beyond the normal range, it may be judged as abnormal speed and initially eliminated.

[0084] The driving time distance

[0085] According to the improved segmented “3σ principle”, the normal range is:

[0086] [D μ.ρi -3D σ.ρi , D μ.ρi +3D σ.ρi ]

[0087] If D ρi If it is beyond the normal area, it may be judged that the driving distance is too different from the planned route or the optimal route, and it may be judged as a detour abnormality and preliminarily eliminated.

[0088] The residence time

[0089] where t pirepresents the time point at which the trajectory point i stops for the pth time, S pi represents the set of time points of the pth stop at trajectory point i, and m represents the total number of stops.

[0090] According to the improved segmented “3σ principle”, the normal range is:

[0091] [T μ.ρstop，pi -3T σ.ρstop，pi , T μ.ρstop，pi +3T σ.ρstop，pi ]

[0092] If T ρstop，pi If it is beyond the normal area, it may be judged as abnormal stay and initially eliminated.

[0093] Among them, the multi-angle trajectory density features include behavior frequency features, time density features, and space density features.

[0094] The behavior frequency f behievr.ρ.i It indicates the number of times a specific behavior occurs when traversing to the i trajectory point in a specific ρ area.

[0095] The time density

[0096] where f time.k.i represents the time density of the i-th trajectory point in the k-th specific time period, T k represents the time point set of vehicles in the area during the kth specific time period, T total，k Represents the total length of the kth specific time period.

[0097] The spatial density

[0098] where f space.ρ.i represents the spatial density of the ith trajectory point in the ρth specific geographic area, N vehicle，ρ represents the number of vehicles in the ρth specific geographical area, A region，ρ represents the area of ​​the ρth specific geographical region.

[0099] S4, weighted fusion of spatiotemporal features, statistical features and trajectory density features into new features for the detection model to learn, so as to detect abnormal vehicle trajectories;

[0100] In step S4, the new feature is a comprehensive feature obtained by fusing multiple features.

[0101]

[0102] Among them, the velocity characteristic (v ρi ), acceleration characteristics (a ρi) and directional characteristics (θ ρi ) is a sub-feature of the spatiotemporal feature, the average velocity feature Driving distance characteristics (D ρi ) and residence time characteristics (T ρstop，pi ) is a sub-feature of the statistical feature; the weights of the internal sub-features in the spatiotemporal feature, statistical feature and spatial density feature are evenly distributed; α, β, γ are the weight coefficients of the spatiotemporal feature, statistical feature and multi-angle trajectory density feature, respectively, and satisfy α+β+γ=1.

[0103] Furthermore, α, β, and γ are weight coefficients, and users can dynamically adjust them by scoring each scenario factor in different scenarios according to their needs, such as Figure 3 As shown;

[0104] S41. Construct feature matrix (H)

[0105]

[0106] Where i is the number of trajectory points, std i represents the standard deviation of each feature corresponding to the i-th trajectory point, so that some highly sensitive features are not underestimated; impi i The importance ranking value of each feature corresponding to the i-th trajectory point is used to determine the importance of each feature, and a larger weight coefficient is assigned to a feature with higher importance, and a smaller weight coefficient is assigned to a feature with lower importance.

[0107] S42. Build scene matrix (G)

[0108]

[0109] Where C represents the scene factor, a represents the number of scene factors, and b represents the number of scenes. In addition, a scene factor C representing the sensitivity to the standard deviation is added. std and scenario factor C, which indicates the importance of ranking imp , which can improve the sensitivity of anomaly detection, adapt to different data distributions, enhance flexibility and customizability, while highlighting key features, reducing noise interference and adapting to different task requirements, thereby optimizing feature weight calculation to improve the accuracy and pertinence of vehicle trajectory anomaly detection.

[0110] S43. Calculate the weight matrix (W)

[0111] First, in order to consider the joint effect of standard deviation and importance ranking, while also considering their respective individual effects, define the function of standard deviation and importance ranking

[0112] f(std, imp)=std*imp+std+imp

[0113] W=H*(G T *diago(f(std1, imp1), f(std2, imp2),, f(std i , imp i )))

[0114] In order to perform personalized weighting on the feature combination of each trajectory point, a diagonal matrix with diagonal elements is added, where diago(f(std1, imp1), f(std2, imp2), f(std i , imp i )) means f(std i , imp i ) is a diagonal matrix with diagonal elements, which allows different trajectory points to be weighted differently according to their standard deviation and importance ranking.

[0115] According to the needs, you can score each scenario factor in different scenarios and obtain dynamic α, β, and γ weight coefficients.

[0116] S5. Training and detection model. Input historical abnormal trajectory data and integrate features into the residual network model. Perform four abnormal trajectory detections on vehicle trajectories, including detour abnormality, speed abnormality, lane change abnormality, and stop abnormality. Define a reward mechanism. Process input features through the residual network and generate trajectory abnormality detection results. According to the detection results and actual results, adjust network parameters through the reward mechanism to optimize detection performance. Use the reward function to evaluate the accuracy of the detection results and obtain the parameter combination with the highest score. After completion, input the vehicle trajectory data set to be identified, detect the trajectory data to be detected through the trained residual network model, and output the abnormality type to complete the abnormal detection of vehicle trajectories.

[0117] In step S5, the historical abnormal trajectory data and fusion features are input into the residual network model, and four abnormal types of trajectory detection are performed on the vehicle trajectory, including: detour abnormality, speed abnormality, lane change abnormality and stop abnormality. A reward mechanism is defined, and the input features are processed through the residual network to obtain preliminary trajectory abnormality detection results. The reward mechanism adjusts the network parameters according to the comparison between the detection results and the actual results, optimizes the detection performance, uses the reward function to evaluate the accuracy of the detection results, and obtains the parameter combination with the highest score of the residual network model. The optimized residual network model is used to detect the trajectory data to be detected, output the abnormality type, and complete the abnormal detection of the vehicle trajectory.

[0118] Among them, S51, input fusion features and historical trajectory data with abnormal type labels into the residual network, design a reward and punishment mechanism, and the reward and punishment score R = R-*n--R +*n +

[0119] Where R- represents the penalty value, which is the penalty value given when the residual network mistakenly detects the vehicle trajectory as normal but is actually abnormal, or mistakenly detects the vehicle trajectory as abnormal but is actually normal or an abnormal type that is not detected; R + Represents the reward value, which is the reward value given when the residual network accurately detects that the vehicle trajectory is normal and is actually normal, or accurately detects that the vehicle trajectory is a certain abnormal type and is actually the abnormal type; n - and n + Represent the number of times of being punished and rewarded respectively.

[0120] Where S52, define the loss function: loss function L

[0121]

[0122] When defining the loss function, not only the difference between the current trajectory data and the actual label is considered, but also the trend of trajectory data changes in the time series. If the trajectory data at multiple consecutive time points show abnormal trends, then the value of the loss function is appropriately increased. represents the actual vehicle trajectory label, Represents the detection result of the residual network, μ is the balance coefficient, n abnormal The number of consecutive abnormal points, τ is the adjustment coefficient. i.l (l∈{0, 1, 2, 3, 4} represents the detection result of the point, including normal and four abnormal types.

[0123] Repeat the above steps to continuously optimize the parameters of the residual network to improve its prediction accuracy. After the residual network is trained, input the trajectory data to be detected, and determine the abnormal trajectory type based on the probability result output by the residual network to complete the abnormal detection of the vehicle trajectory.

[0124] Finally, it should be noted that the above description is only a preferred example of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. 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.

Claims

1. A vehicle trajectory anomaly detection method based on multi-feature fusion, characterized in that: The following steps are involved: S1. Send data to a designated server through the built-in device on the vehicle, then obtain data from the server to obtain vehicle trajectory data; S2, preprocessing the vehicle trajectory data, including data cleaning, data smoothing, noise removal and time alignment; S3, extract trajectory features from the pre-processed vehicle trajectory data set, use the improved segmented 3σ principle to determine the threshold standards of each indicator in the vehicle trajectory data set, mark the vehicle trajectory data outside the threshold as abnormal trajectory data, complete the preliminary screening and elimination of abnormal trajectory data and the marking of abnormal behavior types, and perform further abnormal trajectory detection on the remaining data set; S4, weighted fusion of spatiotemporal features, statistical features and multi-angle trajectory density features into new features for the detection model to learn, so as to detect abnormal vehicle trajectories; S5, training and detection module, inputs historical abnormal vehicle trajectory data and fusion features into the residual network model to perform abnormal trajectory detection on the vehicle trajectory.

2. The vehicle trajectory anomaly detection method based on multi-feature fusion according to claim 1 is characterized in that: In step S1, the vehicle-mounted sensors are used to interact and collect the road condition information of the road section where the vehicle is located in real time, generate detailed lane information and the range, area and number of vehicles in the geographical area, and form relative coordinates according to the current coordinate system and time point and save them together with the road condition information to form a new trajectory data coordinate point (x i ,y i ,t i ,J i ); Among them, the vehicle trajectory data includes global geographic coordinates, vehicle speed and acceleration; Among them, x and y are global relative geographic coordinates, t is the current time point of the trajectory, J represents the road condition of the current trajectory point, and i represents the current trajectory point on the vehicle trajectory.

3. The vehicle trajectory anomaly detection method based on multi-feature fusion according to claim 1 is characterized in that: In step S2: data cleaning, a cleaning method is used to clean the data according to the distribution characteristics of the data, so as to remove duplicate information, blank information and erroneous information in the data set and maintain the uniformity of the characteristics of the data set.

4. The vehicle trajectory anomaly detection method based on multi-feature fusion according to claim 1 is characterized in that: In the step S3, the improved segmented 3σ principle divides the vehicle's driving trajectory according to different road sections, selects vehicle coordinates, collection time and lane information as characteristic indicators, observes the group characteristics of each indicator of the time and space conditions included in each data set, and the group characteristics include spatiotemporal characteristics, statistical characteristics, and multi-angle trajectory density characteristics; and obtains the data intervals where most of the indicators are located for each road section, and the data that does not appear in these intervals shows characteristics different from other data. When a screening indicator in the data record is within the abnormal interval, the data record is marked as abnormal behavior data.

5. The vehicle trajectory anomaly detection method based on multi-feature fusion according to claim 4 is characterized in that: The multi-angle trajectory density features include behavior frequency features, time density features and space density features; Among them, the behavior frequency f behievr.ρ.i It indicates the number of specific abnormal behaviors that occur when traversing to the i trajectory point in a specific ρ area; Among them, the time density Among them, f time.k.i represents the time density of the i-th trajectory point in the k-th specific time period, T k represents the time point set of vehicles in the area during the kth specific time period, T total,k represents the total length of the kth specific time period; Among them, the spatial density Among them, f space.ρ.i represents the spatial density of the ith trajectory point in the ρth specific geographic area, N vehicle,ρ represents the number of vehicles in the ρth specific geographical area, A region,ρ represents the area of ​​the ρth specific geographical region.

6. The vehicle trajectory anomaly detection method based on multi-feature fusion according to claim 1 is characterized in that: In step S4, the new feature is a comprehensive feature obtained by fusing multiple features. Among them, v ρi is the speed characteristic, a ρi is the acceleration characteristic and θ ρi The directional feature is a sub-feature of the spatiotemporal feature. is the average speed characteristic, D ρi is the driving distance characteristic and T ρstop,pi The residence time feature is a sub-feature of the statistical feature; the weights of the internal sub-features in the spatiotemporal feature, statistical feature and spatial density feature are evenly distributed; α, β, and γ are the weight coefficients of the spatiotemporal feature, statistical feature and multi-angle trajectory density feature, respectively, and α+β+γ=1 is satisfied. Users can score each scene factor according to their needs in different scenarios and make dynamic adjustments according to their needs.

7. The vehicle trajectory anomaly detection method based on multi-feature fusion according to claim 6 is characterized in that: α, β, and γ are weight coefficients. Users can score each scenario factor according to their needs in different scenarios and make dynamic adjustments, including: S41. Construct feature matrix H Where i is the number of trajectory points, std i represents the standard deviation of each feature corresponding to the i-th trajectory point, so that some highly sensitive features are not underestimated; impi i The importance ranking value of each feature corresponding to the i-th trajectory point is used to determine the importance of each feature, and a larger weight coefficient is assigned to a feature with higher importance, and a smaller weight coefficient is assigned to a feature with lower importance; S42, construct scene matrix G Where C represents the scene factor, a represents the number of scene factors, and b represents the number of scenes. In addition, a scene factor C representing the sensitivity to the standard deviation is added. std and scenario factor C, which indicates the importance of ranking imp ; S43, calculate weight matrix W First, define the function of standard deviation and importance ranking f(std,imp)=std*imp+std+imp W=H*(G T ·diago(f(std1,imp1),f(std2,imp2),,f(std i ,imp i ))) where diago(f(std1,imp1),f(std2,imp2),,f(std i ,imp i )) means f(std i ,imp i ) is a diagonal matrix with diagonal elements; in different scenarios, score each scenario factor according to needs, and obtain dynamic α, β, and γ weight coefficients respectively.

8. The vehicle trajectory anomaly detection method based on multi-feature fusion according to claim 1 is characterized in that: In the step S5, the trajectory detection of the vehicle trajectory includes four types of anomalies, including detour anomalies, speed anomalies, lane change anomalies and stop anomalies. A reward mechanism is defined, and the input features are processed by the residual network to obtain preliminary trajectory anomaly detection results; the reward mechanism adjusts the network parameters according to the comparison between the detection results and the actual results, optimizes the detection performance, uses the reward function to evaluate the accuracy of the detection results, and obtains the parameter combination with the highest score of the residual network model. The trajectory data to be detected is detected by the optimized residual network model, and the anomaly type is output to complete the abnormal detection of the vehicle trajectory.

9. The vehicle trajectory anomaly detection method based on multi-feature fusion according to claim 8 is characterized in that: In the step S5, S51. Input the fusion features and historical trajectory data with abnormal type labels into the residual network, design a reward and punishment mechanism, and the reward and punishment score R = R - *n - -R + *n + ; Where R - Represents the penalty value, which is the penalty value given when the residual network mistakenly detects the vehicle trajectory as normal but is actually abnormal, or mistakenly detects the vehicle trajectory as abnormal but is actually normal or an undetected abnormal type; R + Represents the reward value, which is the reward value given when the residual network accurately detects the vehicle trajectory as normal and is actually normal, or accurately detects the vehicle trajectory as a certain abnormal type and is actually the abnormal type; n- and n + Represent the number of times of being punished and rewarded respectively; S52. Define the loss function: loss function L Among them, q represents the actual vehicle trajectory label, Represents the detection result of the residual network, μ is the balance coefficient, n abnormal The number of consecutive abnormal points, τ is the adjustment coefficient; Repeat the above steps and continuously optimize the parameters of the residual network until the comprehensive performance of the output results reaches a stable level to improve the prediction accuracy of the model. After the residual network is trained, input the trajectory data to be detected, output the abnormal trajectory type, and complete the abnormal detection of the vehicle trajectory.

Images