A time-series-based dynamic sparse trajectory completion method

Abstract

The application relates to the field of video monitoring track tracking, in particular to a dynamic sparse track completion method based on time sequence. The method comprises the following steps: constructing a mutual mapping model for mutual conversion between a monitoring video and a track vector; acquiring static road features of a target area and abnormal dynamic features based on time sequence; constructing a track tracing model about the target area based on the mutual mapping model and according to the static road features and the abnormal dynamic features of the target area; acquiring monitoring videos of sub-areas in the target area in a demand period and inputting the monitoring videos into the track tracing model, and the track tracing model outputs abnormal dynamic tracks corresponding to the demand period. Through construction of the track tracing model based on the mutual mapping model, the abnormal dynamic tracks corresponding to the demand period about the target can be determined through multi-camera monitoring videos, the action track based on time sequence can be effectively obtained, and thus the problem that an existing track tracing recommendation model cannot be associated with time sequence can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of video monitoring track tracking, and particularly relates to a time sequence-based dynamic sparse track completion method. BACKGROUND

[0002] Most police video monitoring platforms have not realized intelligentization, linkage and integration with geographic information systems, and the value density of data is low, the storage time is short, and data processing cannot be combined with existing track analysis models. The problems have not been effectively solved for a long time. In the big data era, the police department needs to introduce an intelligent track tracing recommendation model to avoid the risk of missing important police information.

[0003] In practice, the historical behavior track of a specific target individual is a sequence set sorted by time, and the data in the set has a shallow or deep dependence relationship. Such sequence information is undoubtedly valuable for track completion. First, it strengthens the influence of recent behavior on subsequent behavior prediction, and second, the sequence model can learn the trend of track preference transition probability. However, the previous track tracing recommendation model does not utilize this sequence information. Even the AFM or DIN model with attention mechanism only scores the importance of different behaviors. Such scores are not related to time or sequence. SUMMARY

[0004] The present application provides a time sequence-based dynamic sparse track completion method, which can solve the problem that the existing track tracing recommendation model cannot be associated with time sequence.

[0005] The technical scheme of the present application is a time sequence-based dynamic sparse track completion method, which comprises:

[0006] S1: constructing a mutual mapping model for mutual conversion between monitoring video and track vector;

[0007] S2: obtaining static road features of a target area and time sequence-based abnormal dynamic features;

[0008] S3: constructing a track tracing model for the target area based on the mutual mapping model and according to the static road features and abnormal dynamic features of the target area; the track tracing model takes the monitoring video of the target area in a target period as an input item and obtains a track vector through the mutual mapping model, and performs trace analysis on the track vector through the static road features and abnormal dynamic features of the target area, to obtain time sequence-based abnormal dynamic track corresponding to the target period, and outputs the abnormal dynamic track;

[0009] S4: obtaining the monitoring video of a sub-area in the target area in a demand period and inputting it into the track tracing model, and the track tracing model outputs abnormal dynamic track corresponding to the demand period.

[0010] Optionally, the step S2 comprises:

[0011] S21: acquiring static road features in the target area, including road density data, congestion situation data, road surrounding POI point data and monitoring point data;

[0012] S22: acquiring a dynamic sparse data set including scattered historical sparse trajectories of a plurality of vehicles in the target area, determining a dynamic trajectory data set including time-based historical connected trajectories of the plurality of vehicles based on a mutual mapping model and according to the dynamic sparse data set, and determining time-based abnormal dynamic features according to the dynamic trajectory data set.

[0013] Optionally, the step S21 comprises:

[0014] S211: performing unit grid division on the target area according to latitude and longitude to obtain a divided area graph, performing coordinate conversion on the divided area graph to obtain a region coordinate conversion graph, performing road density calculation on the region coordinate conversion graph through clustering analysis and performing breakpoint hierarchical analysis on the road density calculation result to obtain road density data in the target area based on different road density intervals;

[0015] S212: obtaining congestion situation data in the target area about expressways, main roads, secondary roads and branch roads respectively in morning peak and evening peak based on a city route traffic congestion identification index;

[0016] S213: obtaining road surrounding POI point data in the target area based on distribution characteristics of administrative units and entertainment places in the target area;

[0017] S214: acquiring monitoring point information in the target area and performing coordinate conversion on the monitoring point information to obtain a monitoring point coordinate conversion graph, and performing kernel density analysis on the monitoring point coordinate conversion graph to obtain monitoring point data including a camera distribution kernel density analysis graph in the target area.

[0018] Optionally, the step S22 comprises:

[0019] S221: acquiring a dynamic sparse data set including scattered historical sparse trajectories of a plurality of vehicles in the target area, and traversing the dynamic sparse data set to obtain all trajectory points of the plurality of vehicles in a scattered form;

[0020] S222: dividing all trajectory points into trajectory point sets corresponding to different vehicles based on numbering bits according to a rule that different vehicles have different numbering bits;

[0021] S223: connecting the trajectory points in the trajectory point sets of different vehicles based on the time sequence by a point set conversion tool to obtain trajectory curves corresponding to the different vehicles, and converting the trajectory curves into trajectory vectors by a mutual mapping model to obtain a dynamic trajectory data set of the historical connected trajectories of the vehicles based on the time sequence;

[0022] S224: periodically mining the dynamic trajectory data set by an unsupervised learning algorithm based on a preset normal behavior distinguishing standard to determine the abnormal dynamic features based on the time sequence in the dynamic trajectory data set.

[0023] Optionally, the trajectory tracing model comprises a feature extraction layer, a trajectory processing layer and an application layer.

[0024] The feature extraction layer is configured to acquire the static road features of the target region and the abnormal dynamic features based on the time sequence.

[0025] The trajectory processing layer is configured to acquire the monitoring video of a sub-region in the target region in a demand period and divide the video monitoring into a plurality of single-frame images, and complete the conversion of the single-frame images into storage nodes by the encoding section of the mutual mapping model to obtain a plurality of storage nodes and trajectory vectors corresponding to the storage nodes.

[0026] The trajectory processing layer is further configured to encode the storage nodes based on the static road features of the target region and the dynamic trajectory behavior, to determine the static road features and the dynamic trajectory features of the storage nodes, and determine the target object in the storage nodes based on the dynamic trajectory features of the storage nodes and confirm the backtracking probability of the actual wandering of the target object.

[0027] The application layer is configured to decode the static road features of the storage nodes and the backtracking probability of the actual wandering of the target object in the storage nodes to obtain the abnormal dynamic trajectory based on the time sequence of the target object corresponding to the demand period, and visually output the abnormal dynamic trajectory corresponding to the demand period.

[0028] Optionally, the step S3 comprises:

[0029] S31: receiving the static road features and the dynamic trajectory features of the target region, learning by the static road features and the dynamic behavior features, to mine the implicit information in the dynamic behavior features and determine the abnormal behavior based on the implicit information, and weighting the abnormal behavior to obtain the trajectory tracing model.

[0030] Optionally, the backtracking probability of the actual wandering of the target object is the probability of observing the target object at the intersection node v i after the last step passes through the adjacent point v i of the intersection node v j , i.e. the node v iBacktracking to the adjacent point v j The calculation formula of the probability is as follows:

[0031] In the formula, ε represents the set of all trajectory edges in the trajectory relationship graph; N + (v i ) represents the set of all in-degree edges of the intersection node v i ; W ij represents the weight of the road backtracking to the intersection node v i . j

[0032] Optionally, the step S4 comprises:

[0033] S41: acquiring monitoring videos of the sub-regions in the target region in the demand period and inputting the monitoring videos to a feature extraction layer of the trajectory tracing model;

[0034] S42: the feature extraction layer separates the monitoring videos into a plurality of single-frame images and respectively adjusts the sizes of the plurality of single-frame images to obtain a plurality of single-frame images with adjusted sizes;

[0035] S43: according to a preset logical mapping relationship and a conversion process rule of a video trajectory space based on the single-frame images to a storage data space, the feature extraction layer converts the single-frame images with adjusted sizes into a visual feature heat map and converts the visual feature heat map into a trajectory curve expressed in a vector form to obtain a storage node;

[0036] S44: according to static road features, the trajectory processing layer establishes a hash index storage table about the storage node, which is used to store the static road features about the sub-regions including road names, road structures, direction bits, end sub-lane bits and target number bits expressed in a vector form;

[0037] S44: according to dynamic trajectory features, the trajectory processing layer determines target objects about the sub-regions in the storage node based on the trajectory curve and confirms the backtracking probability of actual wandering of the target objects;

[0038] S45: the application layer decodes the static road features of the storage node and the backtracking probability of actual wandering of the target objects in the storage node to obtain abnormal dynamic trajectories of the target objects corresponding to the demand period based on time sequences and visually outputs the abnormal dynamic trajectories corresponding to the demand period.

[0039] Optionally, the method for the trajectory processing layer to determine the dynamic trajectory features of the storage node comprises:

[0040] ​The feature extraction layer sequentially performs feature input, feature intersection and feature fusion on the storage nodes based on a graph embedding technology to obtain dynamic trajectory features of the storage nodes

[0041] Advantages:

[0042] The present application can determine the abnormal dynamic trajectory of the target in the demand period through multi-camera monitoring video by constructing a trajectory tracing model based on a mutual mapping model, effectively obtain the time sequence-based action trajectory, and therefore solve the problem that the existing trajectory tracing recommendation model cannot be associated with time sequence.

[0043] In addition, the introduction of the intelligent trajectory recommendation model can realize the deep utilization of video trajectory data, effectively improve the case investigation efficiency of front-line police personnel, reduce police service costs, improve police service quality, and make the trajectory tracing work of the public security department intelligent and automatic, and therefore has important research significance. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0045] Figure 1 The flowchart of a dynamic sparse trajectory completion method based on time sequence in the embodiment of the present application;

[0046] Figure 2 The mutual mapping diagram of the storage node and the intersection node in the embodiment of the present application;

[0047] Figure 3 The VGG11 convolutional neural network structure diagram in the embodiment of the present application;

[0048] Figure 4 The feature extraction comparison diagram after different network tests in the embodiment of the present application;

[0049] Figure 5 The decoding diagram of the embodiment of the present application;

[0050] Figure 6 The data table of the POI point in the embodiment of the present application;

[0051] Figure 7 The schematic diagram before and after the unit grid division of the target area in the embodiment of the present application;

[0052] Figure 8 The natural breakpoint road density grading table in the embodiment of the present application;

[0053] Figure 9 FIG. 14 is a road density analysis chart in an embodiment of the present application;

[0054] Figure 10 FIG. 15 is a chart of early and late peak congestion situation analysis in an embodiment of the present application;

[0055] Figure 11 FIG. 16 is a POI kernel density analysis chart of administrative units and life and entertainment in an embodiment of the present application;

[0056] Figure 12 FIG. 17 is a camera distribution kernel density analysis chart in an urban area in an embodiment of the present application;

[0057] Figure 13 FIG. 18 is an example chart of trajectory points and trajectory lines in an embodiment of the present application;

[0058] Figure 14 FIG. 19 is a structural schematic diagram of a trajectory tracing model in an embodiment of the present application;

[0059] Figure 15 FIG. 20 is a model design diagram of a trajectory tracing model in an embodiment of the present application;

[0060] Figure 16 FIG. 21 is a flowchart of a graph embedding process in an embodiment of the present application;

[0061] Figure 17 FIG. 22 is a comparative schematic diagram of a citizen real trajectory and a trajectory tracing model restored trajectory in an embodiment of the present application;

[0062] Figure 18 FIG. 23 is a comparative schematic diagram of a simulated stolen criminal trajectory and a trajectory tracing model restored trajectory in an embodiment of the present application;

[0063] Figure 19 FIG. 24 is a comparative curve diagram of a trajectory tracing model and other models in an embodiment of the present application. DETAILED DESCRIPTION

[0064] The embodiments will be described in detail below with reference to the drawings. When the following description refers to arrangements in the drawings, identical numbers on different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following embodiments are not meant to represent all implementations consistent with the present application. Rather, they are merely examples of systems and methods consistent with some aspects of the present application as detailed in the appended claims.

[0065] The behavior track of the criminal target in the electric vehicle theft case has obvious space-time characteristics. The path with high road density and weak road congestion situation near the sparse track has a higher accuracy rate of backtracking. Meanwhile, the path with high POI proportion of administrative law enforcement points and dense distribution of law enforcement monitoring positions has a higher probability of not being backtracked. The track backtracking recommendation model can well capture and restore these information and improve the accuracy of track restoration, realize the activity track tracing of the target person, and provide method support for tracing the space-time historical track of the target.

[0066] For the process of historical track completion of a specific target object, the migration of travel preferences is actually very fast. For example, in the first half of the historical track segment, the target object frequently travels according to the shortest route. Therefore, the behavior sequence of the target object in the first half of the journey will be concentrated in the road set of the shortest path. However, after the first half of the journey is completed, the target track object in the second half of the historical track segment starts to frequently avoid congested road segments. Therefore, in the second half of the journey, the travel sequence of the target object will be concentrated in the road set of the average shortest travel time.

[0067] Meanwhile, the expression of the target's hidden information will directly lead to the abnormality of the track. The hidden information can be captured and processed by the autoencoder (Autoencoder). After integrating the hidden information, the surrounding backtracking points of the existing track of the target object will be more interpretable, and the overall track compliance will be higher. In this paper, the theft case is taken as an example to complete the track data completion work. In real public security work, the present application can complete the migration of similar scenes by connecting more relevant data, such as drug and private property seizure work, lost old people patrol work, and suspicious personnel investigation work, etc. The big data technology serves the grassroots security, and makes the grassroots police work more intelligent.

[0068] The present application provides a time sequence-based dynamic sparse track completion method, as shown in Figure 1 The flowchart of the time sequence-based dynamic sparse track completion method in the embodiment of the present application is shown in Figure 1 S1: Construct a mutual mapping model for mutual conversion between monitoring video and track vector.

[0069] S1: Construct a mutual mapping model for mutual conversion between monitoring video and track vector.

[0070] Specifically, the construction of the mutual mapping model includes two processes, namely encoding and decoding.

[0071] Because the image and video data acquired by surveillance cameras occupy a large amount of memory space and have low value density, they are not convenient for long-term storage and utilization. Furthermore, the engineering technology accumulated in the field of computer vision is becoming increasingly mature and has a high accuracy rate for target identification. However, the collaborative calibration work between multiple adjacent cameras is still very difficult to process in the image and video space. Therefore, it is necessary to first perform computer vision processing on the video images and then redefine and organize the identified objects and information so that the data storage and representation format can adapt to the practice of public security tracing.

[0072] When designing the monitor2vec model for the mutual mapping between multi-camera video surveillance images and geospatial information, five important characteristics are defined:

[0073] a) The encoding result should be able to accurately identify any road information, any target's driving direction and trajectory information in the area, while minimizing storage overhead;

[0074] b) The data structure should be simple enough to be combined with multiple models;

[0075] c) This indicates that the results can capture the periodic and non-periodic patterns of the trajectory, and can simulate the target's behavioral habits, preferences, and changes by adjusting the parameters;

[0076] The result represented by d) can be decoded in reverse and the trajectory can be accurately reconstructed in the geographic system;

[0077] e) The embedded recommended encoding result can be interpolated into the decoding sequence.

[0078] Track coding refers to establishing rules for the transformation process from video track space to storage data space through a pre-designed logical mapping relationship in video space, that is, completing the base transformation from video track space to storage data space.

[0079] like Figure 2 As shown, Figure 2 This is a schematic diagram of the mutual mapping between storage nodes and intersection nodes in this application embodiment. The left side is a storage matrix established according to the actual geographical location relationship of intersection nodes. At the index value position corresponding to the matrix node is the address index of the historical monitoring data of the corresponding intersection node stored by the PGIS system. The historical video data is divided into single-frame images and then imported into the deep learning system for feature extraction processing.

[0080] like Figure 3 As shown, Figure 3The VGG11 convolutional neural network structure diagram in the embodiment of the application is taken as an example, and the training set is divided into image frames and imported after adjusting the size. In order to avoid reducing the randomness of training due to an excessively large Bachsize, resulting in that the training result falls into sharp minima, the value is adjusted to 3 to enhance the generalization ability of the model.

[0081] The initial learning rate is set to 0.05, and the learning rate is adjusted in real time using an exponential algorithm. The number of iterations epochs is set to 500. The CAM (Class Activation Mapping) is used to visualize the feature heat map.

[0082] As shown in Figure 4 , Figure 4 The feature extraction comparison diagram after different network tests in the embodiment of the application is shown. The left side is a trajectory diagram, and the right side is a feature heat map, Figure 4 The node-in-depth learning of vehicles at an intersection is performed, and the vehicle flow trajectory point heat map is obtained after machine vision processing. The left side is a video frame whose size is adjusted before being input into each model. The right side is the vehicle flow feature obtained after the VGG (Visual Geometry Group), ResNet (residual network) and ResNext (Split-Attention Networks) respectively, and the CAM (Class Activation Mapping) visual feature heat map of the vehicle flow feature is obtained.

[0083] The continuous vehicle flow point information in the figure is drawn into a trajectory line in the intersection coordinate system, and the trajectory line is expressed in the form of a vector. The road information is basically fixed, and as static data, it is stored and managed separately from the dynamic information such as trajectory.

[0084] At the same time, the road and intersection node information is frequently checked in and checked out when data is called, so a hash index storage table is established for the node of the intersection to store the road name, road structure and other static information.

[0085] For dynamic information such as motion trajectory, the reference also needs to encode the moving direction and sub-lane. The model uses 3 bits to identify 8 direction bits, i.e. east, north, west, south, north to east, north to west, south to west and south to east. The sub-lane information is marked with 3 bits, and the inner lane is numbered from 0 to the outer lane in natural increasing order (since the collected data has a maximum number of 6 from the inner lane to the outer lane), which meets the requirements after testing.

[0086] In summary, for static information such as road names and intersection nodes, the road segment name is entered in 8 bytes, the intersection node information is entered in 2 bytes, 1 bit is reserved for identification, and a hash index table is established respectively, and the index table occupies 1 byte. For dynamic information such as targets and trajectories, 3 bits are used to mark direction information bits (corresponding to eight direction bits), 3 bits are used to mark sub-lane information bits (numbered from the inside lane to the outside lane, and the maximum number is currently 6), 9 bits are used to mark target number bits (the maximum flow statistical value of the intersection in 1 min is <512), and 1 bit is reserved. In this way, only 3 bytes (24 bits) are used to record each trajectory of each intersection.

[0087] As an example of a four-lane road segment in a cross shape, we can uniquely identify a trajectory by road name, node number, direction bit, end sub-lane bit, and target number bit and store it in the database. Due to the establishment of the hash table and the obtained intersection flow data, the estimated value of the amount of trajectory information data generated by each intersection per day is 3*8*7782.5*24 / 1024 / 1024≈4.28G.

[0088] For other types of intersection, the data structure can be reused, and the results are stored in vector form. The periodic and non-periodic pattern mining of trajectories can be calculated by the similarity degree between the encoded vectors.

[0089] For the expression of trajectory features, the only formal definition needed is the backtracking probability of random walk. The actual backtracking probability of the target object is the probability of observing the target object at the intersection node v i after the last step through the adjacent node v i of v j , that is, the calculation formula of the probability of node v i backtracking to adjacent node v j is as follows:

[0090]

[0091] In the formula, ε represents the set of all trajectory edges in the trajectory relationship graph. N + (vi) represents the set of all incoming edges of intersection node v i . W ij represents the weight of the road from intersection node v i backtracking to intersection node v j . The backtracking probability of monitor2vec is the proportion of the weight of the backtracking road in the sum of the weights of the edge set of all related incoming edges.

[0092] At the same time, the results of GraphEmbedding can be made more biased towards the homogeneity or structure of the graph network by adjusting the method of random walk weight.

[0093] Homophily means that the Embedding of nodes with similar distances are as similar as possible, the larger the weight of the in-degree edge, the more the random walk tends to Breadth-First Search (BFS), and the current node is as possible as possible to backtrack to the nearby nodes, while structural means that the Embedding of nodes with similar structures are as similar as possible, the larger the weight of the out-degree, the more the random walk tends to Depth-First Search (DFS), and the current node is more likely to backtrack to the distant nodes.

[0094] The representation form of the trajectory encoding vector provides data consistency for the input calculation and processing of each model. Thus the entire encoding process completes the first three feature requirements of monitor2vec.

[0095] Trajectory decoding is the process of restoring the target behavior trajectory on the grid map after knowing the trajectory encoding or recommended path interpolation encoding, and performing map space labeling.

[0096] As shown in Figure 5 , Figure 5 The decoding schematic diagram of the embodiment of the present application, the dynamic trajectory information encoding 3200 can be identified by the 32 field to locate the static intersection data, and the bidirectional six-lane intersection information is obtained by searching. The target can be known from the 00 field to travel eastward from the starting point through the westbound 0th sub-lane, and the trajectory curve can be easily drawn in the geographic system. Similarly, for other trajectory encodings, the trajectory curve can also be restored by searching the hash table. If part of the intersection has no trajectory encoding data, the trajectory curve can still be easily drawn in the geographic information system according to the trajectory encoding recommended by the post-training model. Thus the decoding process completes the last two feature requirements of monitor2vec.

[0097] In summary, the monitor2vec model completes the five important features defined in the encoding and decoding process, which provides convenience for subsequent application and research, such as importing the vectorized results into the existing deep learning model, and visualizing and analyzing the model recommended results.

[0098] S2: Obtain static road features of a target area and time-based abnormal dynamic features.

[0099] The step S2 includes:

[0100] S21: Obtain static road features in the target area, including road density data, congestion situation data, road surrounding POI point data, and monitoring point data.

[0101] Specifically, the data set required by the embodiment of the application includes theft crime data, road network data, POI point data around the road network, vehicle and personnel trajectory data, and camera information data, etc.

[0102] Theft crime data: Since the proportion of theft crime in criminal cases is high, the overall trajectory of the criminal is greatly affected by the traffic environment and subjective behavior preference, and the behavior characteristics of recidivist personnel are obvious, theft cases are selected for trajectory tracing recommendation model research. The theft case records from 2008 to 2014 in the police receiving platform database of a certain A city branch are used as the theft personnel behavior characteristic data set in the embodiment of the application, and in the subsequent steps, the embodiment will be referred to as the relevant situation of the A city branch, and is specifically used for case interpretation.

[0103] Road network data: The road network data is obtained from the OpenStreet-Map platform. The road traffic line layer of the Chinese region is downloaded, and the layer is processed by using ArcGIS software according to the urban district division to obtain the road network data in the urban district division, and the data coordinates are in the WGS_1984 geographic coordinate system.

[0104] POI point data around the road network: The POI point data around the road network is obtained by using a Python network crawler program to crawl the Gaode map API port. In the embodiment, the data record is 789246, and the POI data attributes include label type, name, address, and longitude and latitude coordinates. In the embodiment, the POI data attributes include administrative units, shopping malls, and life entertainment, and a total of 13249 data. Figure 6 As shown in Figure 6 , the data table of the POI point in the embodiment of the application.

[0105] Vehicle and personnel trajectory data: The vehicle and personnel trajectory data uses the Microsoft T-Driver public data set. In the embodiment, the data set includes the historical trajectory data of 10357 vehicles in the urban area for one week. In addition, the data set includes 15 million coordinate points, and the total distance of the trajectory is more than 9000000 kilometers.

[0106] Camera information data: The camera information data set is also provided by the city branch. In the embodiment, there are 530 law enforcement probes in the city, including camera point latitude and longitude information.

[0107] The step S21 includes:

[0108] S211: Divide the target area into unit grids based on latitude and longitude to obtain a partitioned area map, perform coordinate transformation on the partitioned area map to obtain a regional coordinate transformation map, and calculate the road density on the regional coordinate transformation map through cluster analysis and perform breakpoint hierarchical analysis on the road density calculation results to obtain road density data based on different road density intervals within the target area.

[0109] Specifically, the latitude and longitude span of a certain city's urban area is first divided into grids, dividing the entire area into 100×100 unit grids. Then, the coordinates of the entire urban area, expressways, and grid division zones are transformed, specifically from the WGS_1984 geographic coordinate system to the WGS_1984_UTM_Zone_50N coordinate system. After the transformation, the road density is calculated.

[0110] like Figure 7 As shown, Figure 7 This diagram illustrates the target area before and after grid division in this embodiment of the application. The left side shows the area before division, and the right side shows the area after division. The road lengths within each grid are summed, and the road density of the corresponding grid area is obtained by calculating the ratio of the total road length to the area. Road density is an important indicator for measuring road condition complexity. However, due to the uneven distribution of road density data in urban areas, cluster analysis is performed.

[0111] The natural discontinuity classification method is a commonly used technique in public security practice. By utilizing the characteristics of the data itself, similar objects can be appropriately grouped to maximize the differences between each class. In this paper, the natural discontinuity classification method is used to classify road density.

[0112] Road density elements will be divided into multiple categories, and categories with similar characteristics will be grouped together for further study. Based on the characteristics of urban roads, road density will be segmented into 10 levels.

[0113] The algorithm is based on a small clustering process. The clustering terminates when the variance between groups is maximized and the variance within groups is minimized, and the absolute value of the gradient reaches its minimum, i.e., the GVF value is minimized.

[0114]

[0115]

[0116]

[0117] SDCM i =SDAM 1～i +SDAM i+1～n ;

[0118]

[0119] The break value is calculated by the above formula, wherein p i represents the road density in the grid, the value of which is the ratio of the total length of the road in the grid to the area of the grid, P is the expected value of the overall road density, SDAM is the overall variance, SDCM i is the sum of the inter-group variances of the groups after the selected break point is interrupted, and the sum of the inter-group variances SDCM i takes the minimum value and the overall variance SDAM takes the relatively maximum value, the GVF takes the maximum value of the variance goodness of fit, and the value is between 0 and 1, and the greater the value, the better the fitting. As shown in Figure 8 Figure 8 is a table of natural break road density classification in the embodiment of the present application.

[0120] As can be seen from Figure 8 , the roads in the urban area are divided into ten different classifications according to road density, except that the first-class roads and the tenth-class roads represent the shortest and longest road depths in the grid of each classification, and the road density grouping is mostly concentrated in the third-class roads, the sixth-class roads and the seventh-class roads. After further investigation, the grid road setting is reasonable, and the flow density also shows a similar distribution rule of road density.

[0121] As shown in Figure 9 , the road density analysis diagram in the embodiment of the present application is shown in Figure 9 , which specifically shows the distribution of each level of road after the road density classification, Figure 9 the upper side is each road density interval and the corresponding break value, the middle side is the photographic coordinate system, and the lower side is the column chart corresponding to each road density interval and the corresponding data. It can be seen from Figure 5 that the number of third-class roads in the embodiment of the present application is the largest, which is 1478, followed by the sixth-class roads and the seventh-class roads, which are 709 and 503 respectively, and the roads with the maximum and minimum depths account for 7.89% of the total number of roads. After analysis, it is found that such roads will prolong the time of the activity target staying in the grid, and correspondingly limit the mobility of the activity target in unit time. When the path recommendation is performed subsequently, the corresponding path threshold will be reduced, so that the recommended path avoids falling in the corresponding road set.

[0122] The above analysis results are mainly used to provide a reference basis for the setting of road weights in the subsequent recommendation algorithm, to simulate the general expression of the normal trajectory set in the road environment, and to provide abnormal explanation in trajectory anomaly detection, so it is necessary to analyze.

[0123] S212: Based on the urban line traffic congestion identification index, the congestion situation data of the expressway, the main road, the secondary road and the branch road in the target area during the morning peak and the evening peak is obtained.

[0124] ​Specifically, the road congestion situation is also an influencing factor affecting the historical behavior decision of the target object as a covariate. To quantitatively describe the traffic congestion situation, a traffic congestion recognition index (TCRI) is used to evaluate the running state of the road network. All road segments in the urban area are traversed, and the road segment name and traffic congestion time in the time period are recorded. After being summarized together, the road location information is uniformly corrected in coordinates and then visualized in ArcGIS for analysis.

[0125] In actual operation, the early and late peak congestion situation maps can be obtained, which respectively include: urban early peak traffic congestion situation map and urban late peak traffic congestion situation map. Specifically, the road congestion recognition index identifies the congestion degree of the road segment, and in specific applications, red, yellow and green colors can be used to identify the congestion degree from severe to slight. Thus, it can be seen that the total length and number of congested roads during the early peak and late peak increase significantly, which also affects the occurrence time of theft cases and the selection of criminal target paths.

[0126]

[0127] TCRI, the urban route traffic congestion recognition index, can represent the congestion degree of the rth road segment in the kth period. The larger the value, the more serious the congestion, represents the average driving speed of the ith vehicle on the rth road segment in the kth period, represents the average maximum driving speed of the vehicle on the rth road segment in the kth period.

[0128] As shown in the table, Figure 10 As shown in the table, Figure 10 is an analysis table of the early and late peak congestion situation of the embodiments of the present application, which records the number of sampled road segments, the average speed of the road segment, and the average TCRI, the urban route traffic congestion recognition index, which is obtained by taking the average of the values solved by the above formula. It can be seen that the average speed of the vehicles on the expressway, main and secondary trunk roads changes little during the early and late peak, and the congestion degree of the expressway is similar to that of the main trunk road, while the secondary trunk road is relatively less congested. According to the principle of the shortest time, the travel trajectory will tilt towards the secondary trunk road during the early and late peak.

[0129] S213: Based on the distribution characteristics of administrative units and entertainment places in the target area, obtain the road surrounding POI point data in the target area.

[0130] Specifically, a point of interest (POI) is used to mark important geographic spaces, specifically any meaningful but non-geographical point on a map. This paper studies two types of POI points that have a significant impact on the behavior trajectory of the theft personnel, involving administrative units and life entertainment.

[0131] AsFigure 11 As shown, Figure 11 Fig. 4 is a POI kernel density analysis diagram of administrative units and life and entertainment in an embodiment of the present application. The left side is a POI kernel density analysis diagram of administrative units in the city, and the right side is a POI kernel density analysis diagram of life and entertainment in the city. From Figure 8 It can be seen that the trajectory points of the theft criminal personnel falling on the city road grid will deviate from the city administrative units and move towards the city life and entertainment facilities. The road paths close to the administrative units will be reduced accordingly, and the paths close to the city life and entertainment facilities will be increased.

[0132] S214: Obtain the monitoring point information in the target area and perform coordinate conversion on the monitoring point information to obtain a monitoring point coordinate conversion diagram, and perform kernel density analysis on the monitoring point coordinate conversion diagram to obtain monitoring point position data including a camera distribution kernel density analysis diagram in the target area.

[0133] Specifically, the monitoring point information is derived from the data records of the police station platform database in the city in 2018, with a total of 530 effective records. As a data source of the monitoring point information, the coordinates of the monitoring data points are converted from the Mars coordinate system to the WGS2984 geographic coordinate system for unified processing, and then kernel density analysis is performed.

[0134]

[0135] As shown, Figure 12 As shown, Figure 12 Fig. 5 is a camera distribution kernel density analysis diagram in the city in an embodiment of the present application. The camera dense areas in the city are concentrated in large shopping malls, commercial circles, pedestrian streets and near primary and secondary schools. The camera density aggregation areas falling on the road grid are the entrances and exits of various expressway intersections. In combination with the data of theft cases, it is found that the behavior trajectory of the theft criminal personnel has obvious camera avoidance characteristics. When entering parameters in the trajectory tracing recommendation model, the corresponding road segments will be increased or decreased in weight according to the camera density distribution, so as to highlight the trajectory characteristics of the theft personnel.

[0136] S22: Obtain a dynamic sparse data set of scattered historical sparse trajectories of a plurality of vehicles in a target area, determine a dynamic trajectory data set of historical connected trajectories based on time sequence of the plurality of vehicles based on a mutual mapping model and according to the dynamic sparse data set, and determine an abnormal dynamic feature based on time sequence according to the dynamic trajectory data set.

[0137] Specifically, for the dynamic trajectory data set, the embodiment selects the Microsoft T-Driver public data set to train the model. The data set contains the historical trajectory data of 10357 vehicles in the city in the embodiment for one week, and the data set contains a total of 15 million coordinate points, and the total distance of the trajectories reaches more than 9 million kilometers.

[0138] Step S22 includes:

[0139] S221: Obtain a dynamic sparse dataset containing the scattered historical sparse trajectories of several vehicles in the target area, and traverse the dynamic sparse dataset to obtain all trajectory points of several vehicles distributed in a scattered manner.

[0140] S222: Based on the rule that different vehicles have different numbering positions, all trajectory points are divided into trajectory point sets corresponding to different vehicles based on the numbering positions.

[0141] S223: Based on time series, the trajectory points in the trajectory point sets of different vehicles are connected by a point set to line conversion tool to obtain the trajectory curves corresponding to different vehicles. The trajectory curves are then converted into trajectory vectors through a mutual mapping model to obtain a dynamic trajectory dataset that includes the historical connection trajectories of several vehicles based on time series.

[0142] Specifically, the trajectory datasets of different vehicles in the T-Driver dataset are stored in independent text files. After traversing all vehicle information using Python, the time information and geographic coordinate system latitude and longitude information of all trajectory points are imported into ArcGIS according to the file number. In the trajectory point attribute table, the point set information is distinguished by the number position, that is, the number position of trajectory points of different vehicles is different, and the trajectory point set with the same number position belongs to the same vehicle.

[0143] By using a point-to-line conversion tool, trajectory points with the same vehicle volume are connected in chronological order based on their serial numbers. The entire trajectory line set is then clipped using urban area mapping. Finally, the resulting trajectory curves are processed into trajectory vectors using a mutual mapping model, completing the dataset entry. Figure 13 As shown, Figure 13 This is an example diagram of trajectory points and trajectory lines in an embodiment of this application. The left side of the diagram shows the distribution of trajectory points, and the right side shows the distribution of trajectory lines. The areas where trajectory point lines overlap densely are concentrated on expressways and main roads. In some of these roads, the trajectory lines overlap abnormally, and people are seen taking detours, making them easy to detect.

[0144] S224: Based on a preset normal behavior distinction standard, an unsupervised learning algorithm is used to periodically mine the dynamic trajectory dataset to determine the time-series-based abnormal dynamic features in the dynamic trajectory dataset.

[0145] Specifically, repeat offenders of theft often exhibit behavioral patterns such as loitering near the crime scene before committing theft. Geographic information systems can effectively observe and utilize these patterns.

[0146] After obtaining the trajectory vector, the trajectory is divided into a normal set and an abnormal set by using an unsupervised learning algorithm. After processing by an autoencoder, the normal set usually contains correct driving behavior habits, and the abnormal set contains incorrect driving behavior habits and implicit information, such as deliberately frequent avoidance of monitoring cameras, lane changing, sudden acceleration on a curve, and the like. For the abnormal data set, centralized learning is usually required to mine the implicit information and compare the behavior characteristics of the criminal suspect to further perform targeted trajectory analysis and recommendation, and the trajectory vector of the abnormal behavior is weighted.

[0147] S3: constructing a trajectory tracing model for the target region based on the mutual mapping model and the static road features and the abnormal dynamic features of the target region; the trajectory tracing model takes the monitoring video of the target region in the target period as an input item and obtains a trajectory vector through the mutual mapping model, and performs tracing analysis on the trajectory vector through the static road features and the abnormal dynamic features of the target region, obtains an abnormal dynamic trajectory based on time sequence corresponding to the target period, and outputs the abnormal dynamic trajectory.

[0148] In step S3, the following steps are included:

[0149] S31: receiving the static road features and the dynamic trajectory features of the target region, learning through the static road features and the dynamic behavior features, mining the implicit information in the dynamic behavior features, determining the abnormal behavior based on the implicit information, and weighting the abnormal behavior to obtain the trajectory tracing model.

[0150] As shown in Figure 14 and Figure 15 , the structure of the trajectory tracing model in the embodiment of the present application is shown in Figure 14 Figure 15 The model design diagram of the trajectory tracing model in the embodiment of the present application is shown in

[0151] The feature extraction layer is configured to obtain the static road features of the target region and the abnormal dynamic features based on time sequence.

[0152] The trajectory processing layer is configured to obtain the monitoring video of the sub-region in the target region in the demand period and divide the video monitoring into a plurality of single-frame images, and complete the conversion of the single-frame images to the storage nodes through the encoding section of the mutual mapping model to obtain a plurality of storage nodes and the trajectory vectors corresponding to the storage nodes.

[0153] ​The trajectory processing layer is also used to encode the storage node based on the static road features and dynamic trajectory behavior of the target area, to determine the static road features and dynamic trajectory features of the storage node, and to determine the target object within the storage node based on the dynamic trajectory features of the storage node and to confirm the backtracking probability of the actual movement of the target object.

[0154] The application layer is used to decode the static road features of the storage node and the backtracking probability of the actual movement of the target object within the storage node, to obtain the time-series-based abnormal dynamic trajectory of the target object corresponding to the demand period, and to visualize and output the abnormal dynamic trajectory corresponding to the demand period.

[0155] The methods for determining the dynamic trajectory characteristics of storage nodes in the trajectory processing layer include:

[0156] The feature extraction layer is based on graph embedding technology. It sequentially performs feature input, feature crossing and feature fusion on the storage nodes to obtain the dynamic trajectory features of the storage nodes.

[0157] The trajectory processing layer and the trajectory tracing recommendation model algorithm complete a three-way handshake parameter transmission. The first parameter transmission confirms that the trajectory tracing recommendation model can correctly receive the feature vector from the trajectory processing layer. The second parameter transmission confirms that the trajectory tracing recommendation model algorithm has correctly received the transmitted parameters and can return the recommended trajectory code. The third handshake parameter transmission confirms that the trajectory tracing recommendation model knows that the trajectory processing layer can receive the transmitted parameters. That is, the trajectory processing layer completes reliable data transmission with the trajectory tracing recommendation model through the three-way handshake parameter transmission.

[0158] The application layer then decodes and visualizes the complete path vector returned by the trajectory tracing recommendation model.

[0159] Specifically, in real-world spatial scenarios, the relationship between target objects and data is graph-structured. Objects generate new trajectory sequences through random walks in the trajectory relationship graph. These trajectory sequences are then input into the mutual mapping model to generate embedding vectors.

[0160] like Figure 16 As shown, Figure 16 This is a flowchart illustrating the image embedding process in an embodiment of this application. The left side shows the exported video frames, which are used as time-series images for continuous feature input, then feature crossing is completed in the hidden layer, and finally output after feature fusion.

[0161] The output trajectory vector contains the trajectory information and feature information of the currently observed target. After feature cross-fertilization and fusion of multiple nearby cameras, the feature information can more accurately extract features within the target's focal period, and the better the recommendation accuracy of the model, i.e., the trajectory reconstruction effect.

[0162] an application layer for decoding the static road features of the storage node and the backtracking probability of the actual wandering of the target object within the storage node to obtain the abnormal dynamic trajectory of the target object based on time sequence corresponding to the demand period, and visualizing and outputting the abnormal dynamic trajectory corresponding to the demand period.

[0163] S4: obtaining the monitoring video of the sub-region in the target region in the demand period and inputting the monitoring video into the trajectory tracing model, and the trajectory tracing model outputs the abnormal dynamic trajectory corresponding to the demand period.

[0164] Wherein, the step S4 comprises:

[0165] S41: obtaining the monitoring video of the sub-region in the target region in the demand period and inputting the monitoring video into the feature extraction layer of the trajectory tracing model.

[0166] S42: the feature extraction layer divides the monitoring video into a plurality of single frame images and respectively adjusts the size of the plurality of single frame images to obtain a plurality of single frame images after size adjustment.

[0167] S43: according to the preset logical mapping relationship and the conversion process rule of the video trajectory space based on the single frame image to the storage data space, the feature extraction layer converts the single frame image after size adjustment into a visual feature heat map and converts the visual feature heat map into a trajectory curve expressed in vector form to obtain a storage node.

[0168] S44: according to the static road features, the trajectory processing layer establishes a hash index storage table about the storage node for storing the static road features about the sub-region expressed in vector form including road name, road structure, direction bit, end sub-lane bit and target number bit.

[0169] S44: according to the dynamic trajectory features, the trajectory processing layer determines the target object about the sub-region within the storage node based on the trajectory curve, and confirms the backtracking probability of the actual wandering of the target object.

[0170] S45: the application layer decodes the static road features of the storage node and the backtracking probability of the actual wandering of the target object within the storage node to obtain the abnormal dynamic trajectory of the target object based on time sequence corresponding to the demand period, and visualizes and outputs the abnormal dynamic trajectory corresponding to the demand period.

[0171] Specifically, as shown in Figure 17 , Figure 17For the contrastive schematic diagram of the real trajectory of citizens and the trajectory tracing model restoring trajectory in the embodiments of the present application, the left side is the schematic diagram of the real trajectory of citizens, and the right side is the schematic diagram of the model restored trajectory after sampling. After training by the T-Driver data set, the trajectory tracing model can well simulate and restore the behavior trajectory of ordinary citizens in the focus period, which is basically consistent with the real behavior trajectory.

[0172] As shown in Figure 18 , Figure 18 For the contrastive schematic diagram of the simulated trajectory of theft criminal personnel and the trajectory tracing model restoring trajectory in the embodiments of the present application, the left side is the schematic diagram of the simulated trajectory of theft criminal personnel, and the right side is the schematic diagram of the trajectory tracing model restoring trajectory. For theft criminal personnel, it can be seen that the overall trajectory restoration degree meets the demand of public security tracing work except that the trajectory recommendation of part of the secondary trunk road is deviated. It can be seen that when the trajectory tracing model restores the trajectory, the trajectory of the theft criminal personnel bypassing the check point is accurately simulated at the intersection check point, the hidden information of the target trajectory is successfully captured and expressed, and the path recommendation accuracy for complex road conditions is improved.

[0173] In actual scenarios, the schematic diagram obtained by the trajectory tracing model may have individual outliers with very large deviation degree. Even if the number of outliers is very small, if the root mean square error (RMSE) is used for estimation, the evaluation index will become very poor. In order to make the evaluation process more robust and the evaluation result converge faster, the result is evaluated by using a logarithmic loss function (LogLoss).

[0174] As shown in Figure 19 , Figure 19 For the contrastive schematic diagram of the real trajectory of citizens and the trajectory tracing model restoring trajectory in the embodiments of the present application, the left side is the schematic diagram of the real trajectory of citizens, and the right side is the schematic diagram of the model restored trajectory after sampling. After training by the T-Driver data set, the trajectory tracing model can well simulate and restore the behavior trajectory of ordinary citizens in the focus period, which is basically consistent with the real behavior trajectory. Figure 19 It can be seen that the restoration accuracy of the trajectory tracing model in the embodiments of the present application is compared with the PeriodicMove model, the Seq2Seq model and the TrajGRU model. It can be seen that the model in the present application has improved trajectory restoration accuracy, and the convergence speed of the result is also faster.

[0175] The embodiments of the present application are described in detail above, but the content is only the preferred embodiments of the present application, and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made within the scope of the present application should still belong to the patent coverage range of the present application.

Claims

1. A time-based dynamic sparse trajectory completion method, characterized in that, The method comprises the following steps: S1: constructing a mutual mapping model for mutual conversion between a monitoring video and a trajectory vector; S2: acquiring static road features of a target area and time-based abnormal dynamic features; S3: constructing a trajectory tracing model for the target area based on the mutual mapping model and the static road features and the abnormal dynamic features of the target area; the trajectory tracing model takes a monitoring video of the target area in a target period as an input item, obtains a trajectory vector through the mutual mapping model, and performs tracing analysis on the trajectory vector through the static road features and the abnormal dynamic features of the target area, to obtain time-based abnormal dynamic trajectories corresponding to the target period, and outputs the abnormal dynamic trajectories; S4: acquiring a monitoring video of a sub-area in the target area in a demand period and inputting the monitoring video into the trajectory tracing model, so as to output abnormal dynamic trajectories corresponding to the demand period. The trajectory tracing model comprises a feature extraction layer, a trajectory processing layer and an application layer. The feature extraction layer is configured to acquire the static road features of the target area and the time-based abnormal dynamic features. The trajectory processing layer is configured to acquire the monitoring video of the sub-area in the target area in the demand period, divide the video monitoring into a plurality of single-frame images, and convert the single-frame images into storage nodes through an encoding section of the mutual mapping model, to obtain a plurality of storage nodes and trajectory vectors corresponding to the storage nodes. The trajectory processing layer is further configured to encode the storage nodes based on the static road features and dynamic trajectory behaviors of the target area, to determine the static road features and dynamic trajectory features of the storage nodes, and determine target objects in the storage nodes based on the dynamic trajectory features of the storage nodes and confirm a backtracking probability of actual wandering of the target objects. The application layer is configured to decode the static road features of the storage nodes and the backtracking probability of actual wandering of the target objects in the storage nodes, to obtain time-based abnormal dynamic trajectories of the target objects corresponding to the demand period, and visually output the abnormal dynamic trajectories corresponding to the demand period. The method for determining the dynamic trajectory features of the storage nodes by the trajectory processing layer comprises: The feature extraction layer sequentially performs feature input, feature intersection and feature fusion on the storage nodes based on a graph embedding technology, to obtain the dynamic trajectory features of the storage nodes.

2. The time-based dynamic sparse trajectory completion method according to claim 1, wherein, The step S2 comprises: S21: acquiring static road features in the target area, including road density data, congestion situation data, road surrounding POI point data and monitoring point data; S22: acquiring dynamic sparse data sets of scattered historical sparse trajectories of a plurality of vehicles in the target area, determining a dynamic trajectory data set of time-based historical connected trajectories of the plurality of vehicles based on the mutual mapping model and the dynamic sparse data sets, and determining time-based abnormal dynamic features based on the dynamic trajectory data set.

3. The time-based dynamic sparse trajectory completion method according to claim 2, wherein, The step S21 comprises: S211: The target area is divided into unit grids according to latitude and longitude, a divided area graph is obtained, coordinate conversion is performed on the divided area graph, a region coordinate conversion graph is obtained, road density calculation is performed on the region coordinate conversion graph through cluster analysis, and breakpoint hierarchical analysis is performed on the road density calculation result, to obtain road density data in the target area based on different road density intervals; S212: Based on the urban line traffic congestion recognition index, congestion situation data of the target area about expressways, main roads, secondary roads and branch roads during morning peak and evening peak is obtained; S213: Based on the distribution characteristics of administrative units and entertainment venues in the target area, road surrounding POI point data in the target area is obtained; S214: The monitoring point information in the target area is obtained, coordinate conversion is performed on the monitoring point information, a monitoring point coordinate conversion graph is obtained, and kernel density analysis is performed on the monitoring point coordinate conversion graph, to obtain monitoring point data including a camera distribution kernel density analysis graph in the target area.

4. The method of claim 2, wherein, The step S22 includes: S221: A dynamic sparse data set including scattered historical sparse trajectories of a plurality of vehicles in the target area is obtained, and all trajectory points of the plurality of vehicles in a scattered form are obtained by traversing the dynamic sparse data set; S222: According to the rule that different vehicles have different numbered bits, all trajectory points are divided into trajectory point sets corresponding to different vehicles based on the numbered bits; S223: Based on time sequence, the trajectory points in the trajectory point sets of different vehicles are connected through a point set to line tool to obtain trajectory curves corresponding to different vehicles, and the trajectory curves are converted into trajectory vectors through a mutual mapping model to obtain a dynamic trajectory data set including historical connected trajectory of the plurality of vehicles based on time sequence; S224: Based on a preset normal behavior classification standard, a periodic mining is performed on the dynamic trajectory data set through an unsupervised learning algorithm, to determine abnormal dynamic features based on time sequence in the dynamic trajectory data set.

5. The time-based dynamic sparse trajectory completion method according to claim 1, wherein, The step S3 includes: S31: The static road features and dynamic trajectory features of the target area are received, learning is performed on the static road features and dynamic behavior features, to mine implicit information in the dynamic behavior features and determine abnormal behavior based on the implicit information, and the abnormal behavior is weighted to obtain a trajectory tracing model.

6. The time-based dynamic sparse trajectory completion method according to claim 1, wherein, The backtracking probability of the actual wandering of the target object is that the target object is observed at the intersection node After the occurrence, the last step is passed The adjacent point The probability of the adjacent point The backtracking to the adjacent point The calculation formula of the probability is as follows: ； In the formula, denotes the set of all trajectory edges in the trajectory relation graph; denotes the intersection node the set of all incoming edges; denotes the intersection node back to the intersection node the weight of the road.

7. The time-based dynamic sparse trajectory completion method according to claim 1, wherein, The step S4 includes: S41: The monitoring video of a sub-area in the target area in a demand period is obtained and input to a feature extraction layer of the trajectory tracing model; S42: The feature extraction layer separates the monitoring video into a plurality of single-frame images and respectively adjusts the sizes of the plurality of single-frame images, to obtain a plurality of single-frame images after size adjustment; S43: According to a preset logical mapping relationship and a conversion process rule of a video trajectory space based on a single-frame image to a storage data space, the feature extraction layer converts the single-frame images after size adjustment into a visual feature heat map and converts the visual feature heat map into a trajectory curve expressed in a vector form, to obtain a storage node; S44: According to the static road features, the trajectory processing layer establishes a hash index storage table about the storage node for storing the static road features about the sub-area including the road name, road structure, direction bit, end point sub-lane bit and target number bit expressed in vector form; S44: According to the dynamic trajectory features, the trajectory processing layer determines the target object about the sub-area in the storage node based on the trajectory curve, and confirms the backtracking probability of the actual wandering of the target object; S45: The application layer decodes the static road features of the storage node and the backtracking probability of the actual wandering of the target object in the storage node, obtains the abnormal dynamic trajectory of the target object corresponding to the demand period based on time sequence, and visually outputs the abnormal dynamic trajectory corresponding to the demand period.

Images