A method for identifying vehicle operation behavior types based on sparse trajectories

Abstract

The present invention discloses a method for identifying vehicle operation behavior types based on sparse trajectories. The method steps are as follows: removing invalid data from multi-type vehicle trajectory data through data cleaning to obtain the cleaned vehicle trajectory data; performing data fusion on the cleaned vehicle trajectory data and semantic segmentation data to obtain the fused data; sparsely sampling the trajectory segment data based on monitoring points; training a deep network for travel type classification; inputting a sparse trajectory sequence to be judged, and outputting the driving category distribution of the trajectory segment. The beneficial effects of the present invention are as follows: The present invention can fuse floating car data and multi-source data such as private cars for analysis, providing more insights into vehicle movement behavior patterns and enabling a better understanding of the urban traffic state; it provides the possibility of discovering vehicle movement patterns and offers a new perspective for distinguishing vehicles used for illegal operation.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle trajectory operation, and in particular to a method for identifying vehicle operation behavior types based on sparse trajectories. Background Art

[0002] The development of Internet big data and the popularization of intelligent acquisition devices have greatly simplified the collection of trajectory data of moving objects, and the data resources are becoming increasingly rich. One of the most common types of trajectories is generated by vehicles equipped with driving recorders such as Beidou / GPS. The driving recorder records the position characteristics of the vehicle at a certain time interval to form a section of trajectory. Vehicle trajectory data contains rich spatio-temporal dynamic information, which not only depicts the travel patterns and activity rules of urban residents, but also helps people understand the moving behavior characteristics of different types of vehicles at different stages.

[0003] However, since vehicle trajectory data comes from real life, due to technical limitations or privacy considerations, the trajectory data that can be collected is usually limited; generally, only the driving trajectories of operating vehicles or a few private cars can be collected and recorded; for most vehicles, the traffic management department can usually only obtain the vehicle identification and the time when the vehicle passes through the location of the camera deployed on the road network, and the traffic management department can obtain the sporadic and sparse trajectories of most vehicles.

[0004] 1. As disclosed in a Chinese patent, a method and device for determining a moving trajectory based on sparse trajectory point data (Application No.: CN201310108188.7), by clustering the trajectory point data according to the density of the distribution of trajectory point data on at least two historical trajectories of an object to generate at least two regions, and then determining the regions passed by each historical trajectory according to the at least two historical trajectories and the at least two regions, and determining at least one regional trajectory according to the regions passed by each historical trajectory, so that one regional trajectory can be determined as the moving trajectory of the object according to the at least one regional trajectory and high-frequency trajectory point data, which can avoid the problem in the prior art that the moving trajectory of the object cannot be accurately determined due to the inability to collect enough trajectory point data, thereby improving the reliability of determining the moving trajectory.

[0005] 2. A method for fine classification and recognition of urban road traffic states for sparse trajectory data (Application No.: CN202110842404.5). 1: Collect speed values and calculate the distance of taxi trajectory points relative to the end point of each road section's driving direction as their spatial relative position values. 2: Expand the [speed - space] domain based on the trajectory point's spatial relative position value and speed value, calculate the intersection area of the [speed - space] domains of the front and rear vehicles, and construct vehicle queues for the trajectory points on the road section based on this. Select the optimal queue according to the Davidson - Burgin index. 3: Perform secondary processing on the trajectory queue to obtain the segmentation points for the fine classification of each road section's traffic state. 4: Set the number of traffic state categories, and combine with the "Evaluation Method for Road Traffic Congestion Degree" to obtain the division thresholds for each category of traffic state. Compare the speed values of the vehicle queues in each locally fine - divided road section with the division thresholds of each category of traffic state to obtain the traffic state of each road section. The present invention can achieve fine recognition of urban traffic states.

[0006] Although vehicle trajectory data can be analyzed and recognized in the prior art, the types of vehicle operation behaviors cannot be recognized.

[0007] Therefore, it is necessary to propose a method for recognizing types of vehicle operation behaviors based on sparse trajectories for the above - mentioned problems. Summary of the Invention

[0008] Aiming at the deficiencies in the above - mentioned prior art, the purpose of the present invention is to provide a method for recognizing types of vehicle operation behaviors based on sparse trajectories to solve the above problems.

[0009] A method for recognizing types of vehicle operation behaviors based on sparse trajectories, the method steps are as follows:

[0010] Step 1: Remove invalid data from multi - type vehicle trajectory data through data cleaning to obtain the cleaned vehicle trajectory data, align the operation state data with the cleaned trajectory data to form trajectory segment data marked with operation states.

[0011] Step 2: Perform data fusion on the cleaned vehicle trajectory data and semantic segmentation data to obtain the fused data.

[0012] Step 3: Sparsely sample the trajectory segment data based on monitoring points.

[0013] Step 4: Train a deep network for classifying travel types.

[0014] Step 5: Use the deep neural network obtained by training in Step 4, input the sparse trajectory sequence to be judged, and output the distribution of driving categories of the trajectory segments.

[0015] The further process of Step 1 is as follows:

[0016] (1) Trajectory data cleaning to remove invalid data, redundant data, and drifting data;

[0017] (2) Map matching to match the vehicle position to the corresponding position on the map road and eliminate dynamic drift points;

[0018] (3) Calculate the basic features based on the Beidou / GPS trajectory points.

[0019] Among them, the vehicle trajectory data method in step 2 includes segmenting the trajectory data of operating vehicles and segmenting the trip data of private cars.

[0020] The method for segmenting the trajectory data of operating vehicles is as follows:

[0021] (1) Group all the trajectory data by vehicle ID using Groupby, sort them in ascending order according to the sampling time, and take the set P of trajectory points of a single vehicle and the order data D of this vehicle as the input;

[0022] (2) Take out the departure time StartTime and end time EndTime in the order, and make the following judgment on the sampling time of each trajectory point in the set P. If the time point is between StartTime and EndTime, the current trajectory point is the passenger-carrying trip segment, let Is_occ = 1, otherwise it is the empty-trip segment, let Is_occ = 0;

[0023] (3) Loop steps 1 - 2 until all the trajectory points of all vehicles are processed.

[0024] The process of segmenting the trip of private cars is as follows:

[0025] (1) Similarly, group all the positioning trajectory data by vehicle ID using Groupby and sort them in ascending order according to the sampling time. Take the set P of trajectory points of a single vehicle;

[0026] (2) Starting from the first trajectory point in the set P, it is recorded as the starting point of the first trip of the day;

[0027] (3) Starting from the starting point, rollingly calculate the sampling time interval and the change in speed direction between two adjacent trajectory data points;

[0028] (4) Set the trajectory points with a sampling time interval greater than 15 minutes and a large change in speed direction as the starting points of the next trip segment, and the previous trajectory point as the end point of the previous trip segment.

[0029] Among them, in step 4, constructing a deep network model for sparse trajectory input:

[0030] (1) The multi-attribute embedding layer performs trajectory encoding;

[0031] (2) A recursive module for modeling the trajectory order factor;

[0032] (3) Classify the trajectories using the information extracted by the previous module and assign labels to the trajectories.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows: The vehicle operation behavior type recognition method of the present invention can integrate floating car data and multi-source data such as private cars for analysis, providing more insights into vehicle movement behavior patterns, better grasping the urban traffic status, having certain reference value for further research and policy formulation in urban traffic management. To combat illegal operations, cameras and passing vehicle data at checkpoints are mostly used for manual monitoring. The recognition method provides the possibility of discovering vehicle movement patterns and offers a new perspective for distinguishing vehicles used for illegal operations. By using such sparse trajectories for vehicle behavior analysis and recognition, exploring and mining the spatio-temporal characteristics of sparse data has important practical significance for traffic operation management work such as early warning of abnormal driving behaviors and discovery of illegal operation behaviors. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flowchart of the training process of the vehicle operation behavior classification model based on sparse trajectories of the present invention;

[0035] Figure 2 is a schematic diagram of the sparse sampling method for trajectory segment data based on monitoring points of the present invention;

[0036] Figure 3 is a schematic diagram of the network structure of the sparse trajectory fusion feature classification model based on LSTM of the present invention;

[0037] Figure 4 is a flowchart of the decision-making process of the vehicle operation behavior classification model based on sparse trajectories of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0039] The following will describe the embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention can be implemented in many different ways defined and covered by the claims.

[0040] As Figures 1 to 4 shown, a method for identifying the type of vehicle operation behavior based on sparse trajectories, the method steps are as follows:

[0041] Step 1: Remove invalid data from the multi-type vehicle trajectory data through data cleaning to obtain the cleaned vehicle trajectory data. Align the operation status data with the cleaned trajectory data to form trajectory segment data marked with the operation status. A vehicle trajectory is a time-ordered sequence of the positions visited by a vehicle on a road. Let Traj = <p1p2p3…p i ...p n >, where p i (1 ≤ i ≤ n) is a certain point on the trajectory, determined by the geographical location and timestamp, and can be defined as p i = (lon i , lat i , t i ), where t i is the timestamp, and (lon i , lat i ) is the two-dimensional coordinate composed of longitude and latitude.

[0042] Step 1.1 Trajectory data cleaning

[0043] The trajectory data collected by Beidou / GPS devices includes multiple fields. Usually, fields that are meaningful for trajectory feature analysis, such as instantaneous moment, longitude and latitude, vehicle direction, and instantaneous speed, are retained. At the same time, in combination with the requirements of vehicle movement behavior analysis, observing the vehicle data, the ranges of each field of the collected data are limited according to the corresponding standards for data filtering, which is mainly divided into two parts: Determine the range of the record space position field (longitude and latitude) according to the specific geographical location of the target area. For example, set the range of field Lat to [22.44, 22.79], and the range of field Lon to [113.76, 114.64]. The field Speed records the instantaneous speed of the vehicle at a certain moment and needs to comply with the vehicle speed limit standard on urban roads. Therefore, the range of field Speed is set to [0, 120].

[0044] In addition, when the vehicle parks for a long time and the Beidou / GPS device is still collecting these position points with no obvious change in longitude and latitude, these records are meaningless for analyzing the vehicle movement behavior and can be regarded as redundant data and need to be removed. However, in order to avoid misidentifying traffic conditions (such as traffic jams and traffic lights) as parking points, the standard for identifying a sample point as a long-term parking point is multiple points with adjacent time, unchanged longitude and latitude, and zero speed, and these redundant points are deleted. In this way, meaningful vehicle trajectory points can be obtained and the data storage space can be reduced.

[0045] Due to the influence of hardware devices and environmental factors, etc., the trajectory data collected by Beidou / GPS will not be completely accurate; for example, if the Beidou / GPS device is abnormal, an error signal is generated, resulting in a trajectory drift phenomenon; the drift phenomenon will cause many problems, and the noise point error generated is too large to obtain useful information; Beidou / GPS drift mainly includes static drift and dynamic drift. When the vehicle stops moving but the positioning coordinates are still changing, it is called static drift. Before performing trajectory data mining, these illogical static drift points need to be filtered. The vehicle will also generate Beidou / GPS drift during movement, which is called dynamic drift; for example, a vehicle driving on the road, the next trajectory point positioning deviates from the road, and the trajectory point seriously deviates from the entire trajectory sequence; according to the characteristics of dynamic drift, this paper matches the dynamic drift points to the road network through the map matching method.

[0046] Step 1.2 Map Matching

[0047] Map matching is a key step in studying vehicle driving behavior and identifying network traffic conditions from data. It uses techniques such as probability statistics methods and geometric methods to fuse Beidou / GPS trajectory data and road network information, and matches the position of the trajectory point to the corresponding position on the actual road, that is, matches the longitude and latitude sequence of the vehicle trajectory with the digital map road network; map matching is essentially a pattern matching problem of a plane line segment sequence. According to the analysis of the current research status, the leuven map matching module proposed by Meert and Verbeke is selected to perform map matching preprocessing operations on sparse trajectories.

[0048] Step 1.3 Basic Trajectory Feature Extraction

[0049] In-vehicle Beidou / GPS devices can not only collect position information such as longitude, latitude, and altitude, but also collect the instantaneous speed and direction of the vehicle. The speed in the interval can be obtained through the longitude, latitude, and timestamp information of the Beidou / GPS trajectory points, etc. These are the basic characteristics describing vehicle movement;

[0050] Step 2: Perform data fusion on the cleaned vehicle trajectory data and semantic segmentation data to obtain the fused data. For different means of transportation, there are also differences in the division of travel segments; for subsequent analysis and research, the vehicle travel segments need to be divided according to different methods:

[0051] Step 2.1 Division of Operating Vehicles:

[0052] For operating vehicles, as shown in Table 1, the order data fields record information such as the time and location of passengers getting on and off, and the passenger-carrying information of the vehicle can be judged, and then the travel segments can be divided; first, add a field Is_occ to the operating vehicle trajectory data. When Is_occ is 1, it is a passenger-carrying travel segment, and when it is 0, it is an empty travel segment.

[0053] Table 1 Explanation of Some Fields in Order Data

[0054]

[0055]

[0056] The method for dividing travel segments based on order information is as follows:

[0057] (1) Group all trajectory data by vehicle ID using Groupby and sort them in ascending order according to the sampling time. Take the set of trajectory points P of a single vehicle and the order data D of this vehicle as the input;

[0058] (2) Extract the departure time StartTime and end time EndTime in the order. Make the following judgment on the sampling time of each trajectory point in the set P. If this time point is between StartTime and EndTime, then the current trajectory point is a passenger-carrying travel segment, and set Is_occ = 1; otherwise, it is an empty-load travel segment, and set Is_occ = 0;

[0059] (3) Loop steps 1 - 2 until all trajectory points of all vehicles are processed.

[0060] Step 2.2 Travel Segmenting of Private Car Trajectory Data

[0061] The dataset of private cars has no order data, only trajectory data. Each trajectory point records the vehicle's collection time, location information, and speed information. Select some running characteristics in the trajectory as the reference conditions for finding travel segment points. When a vehicle is driving on the road, except at intersections or in case of emergencies, the vehicle generally does not suddenly change its running direction. The segment point where the movement pattern changes is when the moving object rapidly changes in direction and speed and then tends to be stable. This segmenting algorithm first calculates the changes in speed and direction between a trajectory point and the previous point. If it is greater than the set threshold, it is set as a segment point. However, this method will generate more segment points at traffic lights or in case of traffic jams. Therefore, based on the time interval and changes in speed and direction, the trajectory is segmented. Considering the actual situation, the time interval is set to 15 minutes. If the sampling times of two points on the trajectory differ by more than 15 minutes and the changes in speed and direction are too large, they are respectively set as the end point of the previous travel segment and the start point of the next travel segment. The process of private car travel segmenting is as follows:

[0062] (1) Similarly, group all positioning trajectory data by vehicle ID using Groupby and sort them in ascending order according to the sampling time, with the set of trajectory points P of a single vehicle;

[0063] (2) Start from the first trajectory point in the set P, which is recorded as the starting point of the first travel segment of the day;

[0064] (3) Starting from the starting point, calculate the sampling time interval and the change in velocity direction between two adjacent trajectory data points in a rolling manner;

[0065] (4) Set the trajectory points with a sampling time interval greater than 15 minutes and a large change in velocity direction as the starting point of the next trip segment, and the previous trajectory point as the end point of the previous trip segment.

[0066] Step 2.3 Trajectory segment status annotation

[0067] After dividing the vehicle trip segments, two status labels, empty trip and travel trip, are attached to each trajectory; for vehicle types, including taxis, online passenger transport vehicles, online freight transport vehicles, and private cars, taxis and online car-hailing vehicles have two trip segments, while private cars have only one trip segment; thus, a total of 5 vehicle behaviors can be finally classified, namely, the empty behavior and travel behavior of taxis, the empty behavior and travel behavior of online car-hailing vehicles, and the travel behavior of private cars.

[0068] Step 3: Sparse sampling of trajectory segment data based on monitoring points

[0069] After completing the segmentation and annotation of the trajectory sequence in Step 2, the data of each trajectory segment has the following form: T(traj) = c where traj = <p1, p2, p3,..., p i ,...p N >, p i (1 ≤ i ≤ n) is a certain point on the trajectory, determined by the geographical location and timestamp, and can be defined as p i = (lon i , lat i , t i ), where t i is the timestamp, lon i , lat i is the two-dimensional coordinate composed of longitude and latitude, and C is the travel behavior category;

[0070] c ∈ {taxi empty, taxi travel, online car-hailing empty, online car-hailing travel, private car travel}

[0071] In this step, the trajectory segment data is sparsely sampled based on the road network monitoring points. Let the set of effective monitoring points of the road network be l i = (lon i , lat i , dir i ) ∈ L, where (lon i , lat i ) is the longitude and latitude coordinates of the i-th monitoring point camera, and dir i is the shooting direction of this monitoring camera; based on the traj trajectory and the monitoring point set L, generate the sparse trajectory traj according to the following process stepss : Starting from the first point p1 of the trajectory segment, the monitored points passed through in sequence, i.e., p i ←p1.

[0072] Such as Figure 2 , find in L the monitored point that is close to the trajectory segment (Dis k ≤LW k , where LW k represents the distance proximity threshold, and the value can be selected as the width of the road where the monitored point l k is located), and the monitored point whose camera direction faces the forward direction of the trajectory (|Dir k -Dir(p i+1 , p i )|≤θ Th , where θ Th is the relative direction threshold, and the value can be selected as 15°). If a monitored point l k that meets the conditions is found, then (l k , t i ) will be saved into the sparse trajectory sequence traj s ; p i ←p i+1 . After repeating step (2) and completing the above steps, the sparse trajectory sequence traj s = <(l1, t1), (l2, t2),..., (l n , t n )> and the corresponding travel type T(traj s ) = c are obtained.

[0073] Step 4: Train the deep network for travel type classification.

[0074] Construct a deep network model for sparse trajectory input. Based on the LSTM network structure, input the sparse trajectory trajs and their corresponding types c, and train the network parameters; The LSTM for sparse trajectory classification consists of three parts: (1) The multi-attribute embedding layer encodes the trajectory; (2) The recursive module for modeling the trajectory order factor; (3) Use the information extracted by the previous module to classify the trajectory and assign labels to the trajectory. Such as Figure 3 , mainly including three components.

[0075] The trajectory encoding module encodes the trajectory point features. The longitude and latitude position features of the monitored points are encoded using geohash, and other semantic features around the monitored points, such as surrounding POIs, are encoded using one-hot. Other quantitative features, such as vehicle section speed, number of lanes, driving direction, etc., are directly aggregated with the encoded vector as an extended dimension vector. The time feature of the monitored point is input in the form of mid-week date + time of the day (such as Tuesday, 8:30:00).

[0076] Multiply each attribute by their respective embedding matrices to extract their corresponding embedding representations, and apply an aggregation function to input into the LSTM unit; the LSTM unit captures patterns in the sequence within variable-length time intervals and learns features including the first and last points of the trajectory.

[0077] The output features of the LSTM unit are the hidden features of the trajectory classification module, and each hidden feature output is input into a fully connected layer, with the goal of mapping the learned knowledge to the corresponding labels. The goal of training the model is to minimize the cross-entropy loss of the classification module, as shown in the following formula:

[0078]

[0079] where D train is the training set of trajectory segments, L is the classified trajectory label, C is the travel type, T is the time, and P is the random probability.

[0080] Subsequently, the softmax function is used to emphasize the differences between the labels, and further output the probability distribution of all possible data labels; to avoid model overfitting, dropout and regularization techniques are used; Dropout layers are used throughout the model, so units are randomly discarded during training; in addition, the L1 regularization method is used to regularize the weights and biases of the LSTM unit.

[0081] Step 5: Use the deep neural network obtained by training in Step 4, input the sparse trajectory sequence to be judged, and output the distribution of travel categories of the trajectory segments.

[0082] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structural or equivalent process transformations made using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, are equally included in the patent protection scope of the present invention.

Claims

1. A method for identifying vehicle operation behavior types based on sparse trajectories, characterized in that: The method steps are as follows: Step 1: Remove invalid data from multi-type vehicle trajectory data through data cleaning to obtain the cleaned vehicle trajectory data. Align the operation status data with the cleaned trajectory data to form trajectory segment data marked with the operation status. Step 2: Perform data fusion on the cleaned vehicle trajectory data and the semantic segmentation data to obtain the fused data. Step 3: Sparse sample the trajectory segment data based on monitoring points. Specifically including: 1) Let the set of effective monitoring points of the road network be l i =(lon i , lat i , dir i ) ∈ L, where (lon i , lat i ) is the longitude and latitude coordinates of the camera of the i-th monitoring point, and dir i is the shooting direction of the monitoring camera; based on the traj trajectory and the set of monitoring points L, generate the sparse trajectory traj s according to the following process steps: starting from the first point p1 of the trajectory segment, sequentially pass through the monitoring points, that is, p i ← p1; 2) Search for a distance trajectory segment in L that is close, i.e., satisfies Dis k ≤KW k , where LW k represents the distance proximity threshold, and the value is selected as the width of the road where the monitoring point l k is located, and the monitoring point whose camera direction faces the advancing direction of the trajectory, i.e., satisfies |Dir k -Dir(p i+1 , p i )|≤θ Th , where θ Th is the relative direction threshold. If a monitoring point l k that meets the conditions is found, then (l k , t i ) is saved into the sparse trajectory sequence traj s ; p i ←p i+1 ; 3) Repeat step 2) to obtain the sparse trajectory sequence traj s = <(l1, t1), (l2, t2),..., (l n , t n )> and the corresponding trip type T(traj s ) = c; Step 4: Train a deep network for travel type classification. Step 5: Use the deep neural network obtained by training in Step 4, input the sparse trajectory sequence to be judged, and output the distribution of travel types of the trajectory segments.

2. A method for identifying vehicle operation behavior types based on sparse trajectories as described in claim 1, characterized in that: Wherein the further process of Step 1 is: (1) Trajectory data cleaning, removing invalid data, redundant data, and drift data. (2) Map matching, matching the vehicle position to the corresponding position on the map road to eliminate dynamic drift points. (3) Calculate its basic features according to the Beidou / GPS trajectory points.

3. The vehicle operation behavior type recognition method based on sparse trajectories according to Claim 1, wherein: Wherein the method of the vehicle trajectory data in Step 2 includes segmenting the trajectory data of operating vehicles and segmenting the trips of private car trajectory data.

4. The vehicle operation behavior type recognition method based on sparse trajectories according to claim 3, characterized in that: Wherein the method of segmenting the trajectory data of operating vehicles is as follows: (1) Group all trajectory data by vehicle ID using Groupby and sort them in ascending order according to the sampling time. Use the set P of trajectory points of a single vehicle and the order data D of this vehicle as the input. (2) Take out the departure time StartTime and end time EndTime in the order. Make the following judgment on the sampling time of each trajectory point in the set P. If the sampling time is between StartTime and EndTime, the current trajectory point is a passenger-carrying trip segment, and set Is_occ = 1, otherwise it is an empty trip segment, and set Is_occ = 0. (3) Loop steps 1-2 until all trajectory points of all vehicles are processed.

5. A method for identifying vehicle operation behavior types based on sparse trajectories as described in claim 3, characterized in that: Wherein the process of segmenting private car trips: (1) Similarly group all positioning trajectory data by vehicle ID using Groupby and sort them in ascending order according to the sampling time. Use the set P of trajectory points of a single vehicle. (2) Start from the first trajectory point in the set P and record it as the starting point of the first trip of the day. (3) Starting from the starting point, rollingly calculate the sampling time interval and the change in speed direction between two adjacent trajectory data points. (4) Set the trajectory point with a sampling time interval greater than 15 minutes and a large change in speed direction as the starting point of the next trip, and the previous trajectory point as the end point of the previous trip.

6. A method for identifying vehicle operation behavior types based on sparse trajectories as described in claim 1, characterized in that: Wherein in Step 4, a deep network model for sparse trajectory input is constructed: (1) A multi-attribute embedding layer for trajectory encoding. (2) A recursive module for modeling the trajectory order factor. (3) Use the information extracted by the previous module for trajectory classification and assign labels to the trajectories.

Images