Vehicle motion state estimation method based on traffic millimeter-wave radar

Abstract

This invention discloses a vehicle motion state classification and estimation method based on traffic millimeter-wave radar, mainly addressing the problems of existing methods such as limited application scenarios, low classification and estimation accuracy, high data acquisition sensor costs, and susceptibility to weather conditions. The solution involves: deploying traffic millimeter-wave radar on the roadside to collect trajectory data of vehicles under different motion states, constructing a vehicle motion state dataset; training a support vector machine (SVM) model using the vehicle motion state dataset; classifying vehicle motion states using the trained SVM model; and estimating vehicle motion states by constructing a hidden Markov chain model based on the transition relationships between motion states. The traffic millimeter-wave radar used in this invention is low-cost, adaptable to various adverse weather conditions and traffic scenarios, and significantly improves the accuracy of motion state classification and estimation by constructing a large dataset to train the SVM model. This method can be used for monitoring and early warning of vehicle driving behavior.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology, and further relates to a vehicle motion state estimation method, which can be used to monitor and warn of vehicle driving behavior, thereby improving driving safety. Background Technology

[0002] Currently, intelligent transportation (ITS) is developing rapidly, and traffic safety is its primary goal. Research on vehicle behavior, i.e., the estimation of motion states, plays a crucial role in the development of ITS. The surge in the number of cars on the road has led to increasingly severe traffic accidents and congestion. According to the World Health Organization, approximately 3,400 people die in traffic accidents globally every day, totaling 1.25 million deaths annually. To promote the development of ITS and smart highways and improve the safety of intelligent transportation, diverse IoT sensors are deployed in traffic systems to acquire real-time and accurate vehicle information on the road. However, existing research on vehicle motion state classification and estimation suffers from inadequate classification models and a lack of datasets specifically designed for vehicle motion state classification. This leads to misclassification and misidentification in situations involving complex changes in vehicle motion states, impacting judgments on traffic safety management.

[0003] In their paper "State Transition Mechanism and Classification Algorithm for Static and Dynamic Targets in Front of a Vehicle," published in *Automotive Engineering*, 2016, 38(05):638-645. DOI:10.19562 / j.chinasae.qcgc.2016.05.019, Gao Zhenhai et al. from Jilin University proposed a time window-based method for classifying the state of static and dynamic targets in front of a vehicle. This method classifies motion states into five categories: same direction, opposite direction, start-stop, stationary, and unclassified. Based on these five states, a transition state machine model is established, and then fixed thresholds for different motion states are given empirically. The motion states are then judged using these thresholds. The drawback of this method is that the classification of motion states cannot comprehensively provide information about vehicle driving behavior, such as turning and speeding, and the given thresholds for state transitions lack sufficient data support, resulting in significant limitations in the classification of motion states.

[0004] Shanghai Gaodewei Intelligent Transportation Systems Co., Ltd. disclosed a "Target Vehicle Motion State Recognition Method and Device" in its patent application CN201910879023.7. The method's implementation steps are: determining the relative speed of the target vehicle based on two target images; then obtaining the vehicle's yaw angle based on the rotation angle of the two images; further, obtaining the target vehicle's absolute speed and the position difference between the two images using the relative speed, yaw angle, and vehicle speed information; finally, recognizing the target vehicle's motion state based on its absolute speed and position difference. However, this method uses a video image sensor to collect data, which is susceptible to environmental factors such as weather, leading to inaccurate data, low estimation accuracy, and high hardware costs. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by proposing a vehicle motion state estimation method based on traffic millimeter-wave radar, so as to avoid the influence of weather on data collection, improve the estimation accuracy of vehicle motion state, and reduce hardware costs.

[0006] The technical approach to achieving the objective of this invention is as follows: Vehicle motion data is collected using traffic millimeter-wave radar to avoid the impact of weather on data collection and reduce hardware costs; a vehicle motion state dataset is constructed using the characteristics of vehicle trajectory data under different motion states, and feature values ​​for different motion states are calculated to train a motion state classification model, addressing the problems of insufficient existing data and incomplete motion state classification; considering all possible motion states of a vehicle, motion states are categorized into eight states: normal driving, accelerating, decelerating, vehicle acceleration and start, vehicle deceleration and stop, lane changing, reversing, and stationary; a motion state transition model is constructed to obtain the transition relationships between different motion states; and a hidden Markov chain model for motion state estimation is constructed by calculating the transition probabilities between different motion states to improve the accuracy of motion state estimation.

[0007] Based on the above technical concept, the implementation steps of the present invention include the following:

[0008] (1) Install the traffic millimeter-wave radar on one side of the lane and adjust its installation angle so that its detection range can cover the entire lane;

[0009] (2) Videos of vehicles passing through the millimeter-wave radar monitoring area are recorded by a camera and vehicle trajectory data of vehicles passing through the radar detection area are collected by traffic millimeter-wave radar.

[0010] (3) Construct the vehicle motion state training sample dataset and test sample dataset:

[0011] (3a) From the recorded video and the collected vehicle trajectory data, compare the video time and the timestamp of each frame of data, extract the vehicle trajectory data detected by millimeter-wave radar in each motion state of the vehicle in the video, and record the position, speed and timestamp information of each frame of vehicle trajectory data.

[0012] (3b) Based on the extracted vehicle trajectory data, calculate the velocity characteristic value, acceleration characteristic value and position characteristic value of the vehicle under different motion states;

[0013] (3c) Manually classify the collected vehicle trajectory data based on the calculated velocity characteristic values, acceleration characteristic values, and position characteristic values ​​under different motion states;

[0014] (3d) Label 30% of the classified trajectory data with classification labels to serve as the training sample dataset, and leave 70% of the trajectory data unlabeled to serve as the test sample dataset.

[0015] (4) Use the dataset to train a support vector machine (SVM) model to classify vehicle motion states:

[0016] (4a) Let M be the number of all motion state categories of the vehicle. Design a binary classification support vector machine (SVM) between any two motion state categories. This requires... A binary classification support vector machine (SVM);

[0017] (4b) Train all binary classification support vector machine (SVM) models using the training sample dataset:

[0018] (4c) Input the test dataset T into each binary support vector machine (SVM) model to obtain the classification result of each binary model;

[0019] (4d) Perform a vote on the classification results of all binary classification models, and the result with the most votes is the classification result of the test data.

[0020] (5) Establish a hidden Markov chain model λ for vehicle motion state estimation:

[0021] (5a) Define the set of all possible motion states as Q = {q1, q2, q3} i …,q N The set of all possible observations is O = {o1, o2, o}. j …,o M The motion state sequence S of the trajectory data for which motion state estimation is required is given by the observation sequence V, where S = {s1, s2, s3}. t …,s T}, V={v1,v2,v t …,v T}, where qi For the i-th possible motion state, o j Let s be the j-th possible observed motion state, N be the number of possible hidden states, M be the number of possible observed states, and s be the number of observed motion states. t Let v be the state sequence of frame t. t Let t be the observation sequence of the t-th frame, T be the sequence length, and s be any hidden state. t ∈Q, any observed state v t ∈O;

[0022] (5b) Based on the eight possible motion states of a vehicle, namely normal driving, accelerating driving, decelerating driving, vehicle acceleration and starting, vehicle deceleration and stopping, lane changing driving, reversing driving and standing still, determine the transition relationship between these states and establish the state machine of the vehicle's motion state.

[0023] (5c) Based on the statistical results of classifying a large amount of vehicle data in step 4 and the transition relationship between each motion state, calculate the state transition probability a of the hidden Markov chain. ij Observation state generation probability b j (k) and initial state distribution π:

[0024]

[0025] M is the number of vehicle motion state categories, M = 8;

[0026] π = [π(i)] N ,π(i)=P(s1=q i )

[0027] (5d) Based on the state transition probability a ij and the probability of generating the observed state b j (k) are used to form the state transition matrix A and the observation state generation probability matrix B, respectively, to construct the hidden Markov chain model: λ=(A,B,π);

[0028] (6) Based on the hidden Markov chain model established in step (5) and the motion state classification results obtained in step (4), the Viterbi algorithm is used to calculate the state q at time t. i The probability of a state is the probability of the state. The state with the highest probability among the probabilities of N states is the estimated state.

[0029] (7) Transmit the estimation results obtained in (6) and the classification results obtained in (4) to the intelligent transportation cloud platform.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] First, because the present invention uses traffic millimeter-wave radar deployed on the roadside to collect vehicle trajectory data, it overcomes the problems of existing technologies that use cameras to collect data, which are easily affected by weather conditions and have high hardware costs, as well as the lack of uniformity in vehicle radar systems and incomplete information utilization. This invention enables the present invention to obtain accurate vehicle trajectory data in any weather conditions.

[0032] Secondly, since this invention considers all eight possible motion states of vehicles and constructs a transformation model between motion states, it overcomes the problem in the prior art that the classification information of vehicle motion states is incomplete, resulting in incomplete traffic information provided to the traffic system. This invention can provide accurate and complete vehicle driving information to the intelligent transportation system and provide accurate and effective services for traffic safety management.

[0033] Third, because this invention collects a large amount of vehicle trajectory data under various motion states on the road, statistically analyzes vehicle driving characteristics under different motion states, and constructs a vehicle motion state dataset, it overcomes the problem of insufficient test datasets in existing technical solutions, which leads to limitations in the application scenarios of the solutions. This makes the invention applicable to any traffic scenario for classifying and recognizing vehicle motion states, while greatly improving the estimation accuracy of motion states. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0035] Figure 2 This is a diagram of the vehicle motion state transition state machine model in this invention; Detailed Implementation

[0036] The embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] Reference Figure 1 The implementation steps of this embodiment are as follows:

[0038] Step 1: Collect vehicle trajectory data under different motion states.

[0039] A traffic millimeter-wave radar is installed on one side of a road with complex traffic conditions. The radar is mounted on a bracket 3.5 meters high, and the radar angle is adjusted so that it can detect the driving trajectory data of vehicles passing by within the maximum range.

[0040] The traffic millimeter-wave radar used in this embodiment is the Huiershi integrated radar-visual millimeter-wave radar, but it is not limited to other traffic millimeter-wave radars on the market that can detect vehicle driving data. This radar integrates a radar module and a video module. Each millimeter-wave radar can simultaneously collect millimeter-wave radar data and video data of vehicles passing by. The acquired millimeter-wave radar data includes vehicle speed, position, and acceleration information.

[0041] Step 2: Establish a motion state transition state machine model.

[0042] 2.1) The possible driving states of a vehicle during road travel are divided into eight states: normal driving, accelerating driving, decelerating driving, vehicle acceleration and starting, vehicle deceleration and stopping, lane changing driving, reversing driving, and stationary driving.

[0043] 2.2) Based on the eight motion states of the vehicle, establish the following three transition relationships between motion states:

[0044] The first type: That is, from a stationary state to the vehicle acceleration and start state, from the vehicle acceleration and start state to the normal driving state, from the normal driving state to the acceleration driving state, from the acceleration driving state to the normal driving state or the deceleration driving state, from the deceleration driving state to the acceleration or normal driving state, from the normal driving state to the deceleration or lane changing state, from the lane changing state to the normal driving state or the vehicle deceleration and stop state, and from the vehicle deceleration and stop state to the stationary state.

[0045] The second type: That is, from a stationary state to the vehicle acceleration and start state, from the vehicle acceleration and start state to the normal driving state, from the normal driving state to the acceleration driving state, from the acceleration driving state to the normal driving state or deceleration driving state, from the deceleration driving state to the acceleration or normal driving state, from the normal driving state to the deceleration driving state or the vehicle deceleration and stop state, and from the vehicle deceleration and stop state to the stationary state.

[0046] The third type: stationary → vehicle acceleration → reversing → vehicle deceleration → stationary, that is, from stationary state to vehicle acceleration state, from acceleration state to reversing state, from reversing state to vehicle deceleration state, and from vehicle deceleration state to stationary state.

[0047] 2.3) Draw the motion state transition model based on the three transition relationships between the eight motion states, i.e., the state machine. Figure 2 As shown.

[0048] Step 3: Construct a dataset of vehicle motion states.

[0049] 3.1) Save the trajectory data of vehicles passing by the traffic millimeter-wave radar, compare the video recorded during data collection with the timestamp of the saved data, extract the trajectory data corresponding to the actual movement state changes of the vehicle, and record the position, speed and timestamp information of the vehicle.

[0050] 3.2) After the traffic millimeter-wave radar collects the trajectory data of the vehicle in all states, the collected trajectory data is analyzed and processed to calculate the velocity characteristic value, acceleration characteristic value, and position characteristic value of the vehicle in different motion states:

[0051] 3.2.1) Calculate the velocity characteristic values ​​under different motion states:

[0052] The average speed v of the vehicle was calculated using the sliding window method with X-axis coordinate information for every 10 consecutive frames of data. x The velocity feature value v in the X-axis direction of the i-th frame of data x i :

[0053]

[0054] The average speed v of the vehicle was calculated using the sliding window method with Y-axis coordinate information for every 10 consecutive frames of data. y The velocity feature value v in the Y-axis direction of the i-th frame of data y i :

[0055]

[0056] Where, x i The x-axis coordinate information of the i-th frame of data, y i For the Y-axis coordinate information of the i-th frame of data, t i v is the timestamp of the i-th frame of data. x i Let v be the velocity feature value in the X-axis direction of the i-th frame of data. y i The velocity feature value in the Y-axis direction of the i-th frame of data;

[0057] 3.2.2) Calculate the characteristic values ​​of acceleration under different motion states:

[0058] Based on the X-axis velocity characteristic value obtained above, the average X-axis acceleration of the vehicle for every 10 consecutive frames is calculated using the sliding window method, and this is taken as the acceleration characteristic value a in the X-axis direction of the i-th frame. x i :

[0059]

[0060] Based on the velocity characteristic value in the Y-axis direction obtained above, the average Y-axis acceleration of the vehicle for every 10 consecutive frames is calculated using the sliding window method, and this is taken as the acceleration characteristic value a in the Y-axis direction of the i-th frame. y i :

[0061]

[0062] Among them, v x i The velocity information of the i-th frame of data in the X-axis direction, v y i For the velocity information of the i-th frame of data in the Y-axis direction, t i Let a be the timestamp of the i-th frame of data. x i Let a be the acceleration characteristic value of the i-th frame of data in the X-axis direction. y i Let be the acceleration characteristic value of the i-th frame of data in the Y-axis direction;

[0063] 3.2.3) Calculate the positional characteristic values ​​under different motion states:

[0064] The position information of each frame of data on a trajectory is determined, and the first 15 frames of data for each trajectory are defined as the starting stage data, and the data after the 15th frame are defined as the ending stage data. Then, the sliding window method is used to calculate the change in the Y-axis coordinate of the vehicle every 15 consecutive frames: d y i =y i+15 -y i ,

[0065] Where, d y i Let y be the change value of the i-th frame of data in the Y-axis direction. i Let be the coordinate value of the i-th frame of data along the Y-axis.

[0066] 3.3) Classify the trajectory data based on the vehicle's speed, acceleration, and position characteristics under different motion states;

[0067] 3.3.1) Let v x d The velocity threshold in the X-axis direction, a x d The acceleration threshold in the X-axis direction, d y d The threshold for position change in the Y-axis direction, v y d The velocity threshold in the Y-axis direction;

[0068] 3.3.2) Based on the characteristic values ​​of the vehicle under different motion states and the threshold parameters set in (3.3.1), determine the different states of the vehicle trajectory data:

[0069] If the trajectory data simultaneously satisfy: v x i ≤v x d a x i ≤-a x d If the vehicle is in a deceleration and stopping state at that moment, and the trajectory data is at the end of the trajectory, then the vehicle is in a deceleration and stopping state at that moment.

[0070] If the trajectory data simultaneously satisfy: v x i ≤v x d a x i ≥a x d If the trajectory data is in the initial stage of the trajectory, then the vehicle is in an acceleration and start-up state at that moment.

[0071] If the trajectory data simultaneously satisfy: a x i ≥a x d If the trajectory data is at the end of the trajectory, then the vehicle is accelerating at that moment.

[0072] If the trajectory data simultaneously satisfy: a x i ≤-a x d If the trajectory data is at the end of the trajectory, then the vehicle is in a state of deceleration at that moment.

[0073] If the acceleration characteristic value of the trajectory data is in [-a] x d ,a x d If the value is between ], then the vehicle is in normal driving condition at that moment;

[0074] If the trajectory data simultaneously satisfy: v y i ≥v y d d y i ≥d y d If so, the vehicle is in a lane-changing state at that moment;

[0075] If the velocity characteristic value of the trajectory data in the X-axis direction is negative, then the vehicle is in a reversing or reverse driving state at that moment; where v x i Let v be the velocity feature value in the X-axis direction of the i-th frame of data. y i Let a be the velocity feature value in the Y-axis direction of the i-th frame of data. x i Let a be the acceleration characteristic value of the i-th frame of data in the X-axis direction. y i Let d be the acceleration characteristic value of the i-th frame of data in the Y-axis direction. y i This represents the change in the data of the i-th frame along the Y-axis.

[0076] 3.4) Label 30% of the trajectory data obtained under different motion states and use it as the training sample dataset X, and use 70% of the unlabeled trajectory data as the test sample dataset T.

[0077] Step 4: Use the dataset to train a support vector machine (SVM) model to classify the vehicle's motion state.

[0078] 4.1) Let M be the number of all motion state categories of the vehicle. Design a binary classification support vector machine (SVM) between any two motion state categories. This requires... A binary classification support vector machine (SVM);

[0079] 4.2) Train each binary classification support vector machine (SVM) model using the training sample dataset:

[0080] 4.2.1) Convert the training sample dataset X into a vector, i.e., X = {(x1, y1), ..., (x... i ,y i ),...,(x N ,y N}, x i For data features, y i For data labels;

[0081] 4.2.2) Call the libsvm toolkit, input the training sample data X, and automatically retrieve the optimal values ​​of parameters gamma and c. The parameters gamma and c represent the performance and fault tolerance of the support vector machine (SVM), respectively.

[0082] 4.2.3) After adjusting the parameters, set the optimal values ​​of gamma and c, and then call the svm-train program in the libsvm toolkit to obtain a trained binary classification support vector machine (SVM) model.

[0083] 4.3) Repeat step (4.2) to obtain all binary classification support vector machine models;

[0084] 4.4) Input the test dataset T into each binary classification support vector machine (SVM) model to obtain the classification result of each binary classification model;

[0085] 4.4) Perform a vote on the classification results of all binary classification models, and the result with the most votes is the classification result of the test data.

[0086] Step 5: Establish the Hidden Markov Chain Model.

[0087] 5.1) Set the parameters required for model calculation:

[0088] Define the set of all possible motion states as Q = {q1, q2, q...} i …,q N}, where q i Let N be the i-th possible motion state, and N be the number of possible hidden states.

[0089] Define the set of all possible observations as O = {o1, o2, o3} j …,o M}, where o j It is the j-th possible observed motion state, and M is the number of possible observed states;

[0090] Define the motion state sequence of the trajectory data for which motion state estimation is required as S = {s1, s2, s3}. t …,s T}, where s t Let be the state sequence of frame t, where T is the sequence length and s is any value. t ∈Q;

[0091] Define the observation sequence of trajectory data for which motion state estimation is required as V = {v1, v2, v...} t …,v T}, where v t Let v be the observation sequence of the t-th frame, where T is the sequence length, and v is any value. t ∈O;

[0092] 5.2) Based on the statistical analysis of the classification results of a large amount of vehicle data in step 4, calculate the state transition probability α of the hidden Markov chain model. ij Observation state generation probability b j (k) and initial state distribution π:

[0093]

[0094]

[0095] π = [π(i)] N ,π(i)=P(s1=qi )

[0096] 5.3) Construct a state transition matrix A based on the calculated transition probabilities of all motion states.

[0097]

[0098] 5.4) Based on the calculated probabilities of the observed states for all motion states, construct the observed state generation probability matrix B.

[0099]

[0100] 5.5) Based on the state transition matrix A, the observed state generation probability matrix B, and the initial probability distribution π, the HMM model is obtained, denoted by λ as follows:

[0101] λ = (A, B, π).

[0102] Step 6: Use the Viterbi algorithm to estimate the vehicle's motion state.

[0103] 6.1) Based on the Hidden Markov Chain Model established in step 5, use the Viterbi algorithm to calculate the state q at time t. i The probability of:

[0104] q i =P(i t =i,i t-1 ,...,i1,o t ,...,o1|λ)=δ t-1 (j)a ij ]b i (o t ), i = 1, 2, ..., N

[0105] 6.2) Selecting the state with the highest probability from the probabilities of N states is the estimated state δ. t (i):

[0106]

[0107] Where, δ t-1 (j) represents the probability of the state at the previous time step, a ij Let b be the state transition probability from state i to state j. i (o t Let be the observation probability of state i at time t.

[0108] Step 7: The traffic system will upload the vehicle driving status classification information obtained in Step 4 and the estimation information obtained in Step 6 to the intelligent transportation cloud platform.

[0109] The effectiveness of this invention can be further illustrated by the following test data:

[0110] I. Test Environment and Conditions

[0111] Huier Vision traffic millimeter-wave radar was used to collect vehicle trajectory data under various motion states at the Zhangba East Road Dongkou bus stop in Yanta District, Xi'an. MATLAB software was used to analyze and test the collected vehicle trajectory data.

[0112] The test sample dataset in the vehicle motion state classification dataset constructed for testing contains more than 6,000 sample data of vehicles in five motion states: acceleration start, acceleration driving, normal driving, deceleration driving, and deceleration stopping.

[0113] II. Test Content and Results

[0114] Test 1, under the above test environment and conditions, performed a classification test on more than 6000 sample data. The performance evaluation is shown in Table 1:

[0115] Table 1. Classification and performance evaluation of vehicles under different motion states

[0116]

[0117] As shown in Table 1, the comprehensive evaluation of the vehicle's classification accuracy for the five states—acceleration start, acceleration driving, normal driving, deceleration driving, and deceleration stopping—was 84.44%, 98.99%, 97.36%, 96.93%, and 88.89%, respectively, with an overall classification accuracy of 97.11%.

[0118] Test 2: Under the above test environment and conditions, the data of more than 6,000 sample datasets of the test sample dataset were estimated. That is, the estimated vehicle motion state of the test sample dataset was compared with the actual vehicle motion state of the test sample dataset. The vehicle motion state estimation accuracy was 98.66%.

[0119] The above is merely a specific embodiment of the present invention and does not constitute any limitation on the present invention. Obviously, those skilled in the art can make various modifications and changes in form and detail under the idea and spirit of the present invention, but these modifications and changes based on the idea of ​​the present invention are still within the protection scope of the claims of the present invention.

Claims

1. A traffic millimeter wave radar-based vehicle motion state classification and estimation method, characterized by, Comprise as follows: (1) install traffic millimeter wave radar on one side of the lane, and adjust the installation angle to make its detection range cover the entire lane; (2) record the video of vehicles passing through the millimeter wave radar monitoring area by the camera and collect vehicle trajectory data of vehicles passing through the radar detection area by the traffic millimeter wave radar; (3) construct vehicle motion state training sample data set and test sample data set: (3a) from the recorded video and collected vehicle trajectory data, compare the video time and the time stamp of each frame data, extract the vehicle trajectory data detected by the millimeter wave radar in each motion state of the vehicle in the video, and record the position, speed and time stamp information of each frame of vehicle trajectory data; (3b) according to the extracted vehicle trajectory data information, calculate the speed characteristic value, acceleration characteristic value and position characteristic value of the vehicle in different motion states; (3c) according to the calculated speed characteristic value, acceleration characteristic value and position characteristic value in different motion states, manually classify the collected vehicle trajectory data; (3d) label 30% of the classified trajectory data with classification labels as the training sample data set, and 70% of the trajectory data without classification labels as the test sample data set; (4) use the data set to train support vector machine SVM model to classify the vehicle motion state: (4a) Let the number of all motion state categories in which the vehicle moves be M, and a binary classification support vector machine SVM is designed between any two motion state categories, a total of binary classification support vector machines SVMs are needed. (4b) use the training sample data set to train all binary support vector machine SVM models: (4c) input the test data set T into each trained binary support vector machine SVM model to obtain the classification results of each binary model; (4d) vote for the classification results of all binary models, and the result with the most votes is the classification result of the test data; (5) establish vehicle motion state estimation hidden Markov chain model λ: (5a) Define all possible motion state sets as Q = {q1, q2, q i … N}, all possible observation sets as O = {o1, o2, o j … M}, the motion state sequence S of the trajectory data that needs to be estimated as S = {s1, s2, s t … T} and the observation sequence V as V = {v1, v2, v t … T}, where q i is the i-th possible motion state, o j is the j-th possible observation state, N is the number of possible hidden states, M is the number of possible observation states, s t is the t-th frame state sequence, v t is the t-th frame observation sequence, T is the sequence length, any hidden state s t ∈ Q, and any observation state v t ∈ O; (5b) according to the 8 possible motion states of the vehicle, i.e. normal driving, accelerating driving, decelerating driving, vehicle accelerating start, vehicle decelerating stop, lane changing driving, reversing driving and static, determine the transition relationship between the states, and establish the state machine of vehicle motion state; (5c) According to the statistics of the classification results of the mass vehicle data and the transition relationship between each motion state, the state transition probability a of the hidden Markov chain is calculated respectively ij , the observation state generation probability b j (k) and the initial state distribution π: M is the number of vehicle motion state categories, M = 8; π = [π(i)] N , π(i) = P(si = q i ) (5d) According to the state transition probability a ij and the observation state generation probability b j (k), respectively, constitute the state transition matrix A and the observation state generation probability matrix B, and the hidden Markov chain model is constructed: λ = (A, B, π); (6) Using the hidden Markov chain model established according to step (5) and the motion state classification result obtained in step (4), the probability of state q i at time t is calculated by using the Viterbi algorithm, and the state with the maximum probability is selected from the probabilities of N states, which is the estimated state; (7) transmit the estimation results obtained in (6) and the classification results obtained in (4) to the intelligent traffic cloud platform.

2. The method of claim 1, wherein, In step (3b), the speed characteristic value, acceleration characteristic value and position characteristic value of the vehicle in different motion states are calculated as follows: (3b1) Using the coordinate information of the X-axis, the average speed v of the vehicle running every 10 continuous frames of data is calculated using the sliding window method x As the speed feature of the X-axis direction of the current frame, the average speed v of the vehicle running every 10 continuous frames of data is calculated using the coordinate information of the Y-axis using the sliding window method y As the speed feature of the Y-axis direction of the current frame: wherein x i is the X-axis coordinate information of the i-th frame data, y i is the Y-axis coordinate information of the i-th frame data, t i is the time stamp of the i-th frame data, v x i is the speed feature value of the i-th frame data in the X-axis direction, v y i is the speed feature value of the i-th frame data in the Y-axis direction; (3b2) Using the speed information calculated in (3b1), calculate the average acceleration a of the vehicle for every 10 consecutive frames using the sliding window method x and a y Acceleration features in the X and Y axes directions as the current frame: wherein, v x i is the speed information of the i-th frame data in the X-axis direction, v y i is the speed information of the i-th frame data in the Y-axis direction, t i is the timestamp of the i-th frame data, a x i is the characteristic value of the acceleration of the i-th frame data in the X-axis direction, a y i is the characteristic value of the acceleration of the i-th frame data in the Y-axis direction; (3b3) determine the position information of each frame data of a track, define the first 15 frame data of each track data as the start stage data of a track, define the data after 15 frames as the end stage data of a track, calculate the change value d of Y axis coordinate of each continuous 15 frames of vehicle driving by using sliding window method y : d y i = y i+15 - y i wherein d y i is the change value of the i-th frame data in the Y-axis direction, y i is the coordinate value of the i-th frame data in the Y-axis direction.

3. The method of claim 1, wherein, In step (3c), the trajectory data is classified by speed characteristics, acceleration characteristics and position characteristics as follows: (3c1) Set v x d is a velocity threshold in the X-axis direction, a x d is an acceleration threshold in the X-axis direction, d y d is a position change threshold in the Y-axis direction, v y d is a velocity threshold in the Y-axis direction; (3c2) according to the characteristic values of the vehicle in different motion states and the threshold parameters set in (3c1), determine the different states of the vehicle trajectory data: If the trajectory data satisfies the following three conditions simultaneously: v x i ≤v x d , a x i ≤-a x d , and the trajectory data is in the end stage of the trajectory, the vehicle is in the deceleration and stop state at this moment. If the trajectory data satisfies: v x i ≤v x d , a x i ≥a x d , and the trajectory data is in the trajectory starting stage, the vehicle is in the accelerating starting state at this moment. If the trajectory data satisfies: a x i ≥ a x d and the trajectory data is in the end stage of the trajectory, the vehicle is in an accelerating driving state at this moment. If the trajectory data satisfies: a x i ≤ -a x d , and the trajectory data is in the end stage of the trajectory, the vehicle is in a deceleration driving state at this moment. If the acceleration eigenvalue of the trajectory data is between [-a x d ,a x d ], the vehicle is in normal driving state at this moment. If the trajectory data satisfies the following conditions simultaneously: v y i ≥ v y d , d y i ≥ d y d , the vehicle is in a lane-changing state at this moment. If the trajectory data is negative in the X-axis direction speed characteristic value, then the vehicle is in the reversing state at this moment; wherein v x i is the speed feature value of the i-th frame data in the X-axis direction, v y i is the speed feature value of the i-th frame data in the Y-axis direction, a x i is the acceleration feature value of the i-th frame data in the X-axis direction, a y i is the acceleration feature value of the i-th frame data in the Y-axis direction, d y i is the change value of the i-th frame data in the Y-axis direction.

4. The method of claim 1, wherein, In step (4b), all binary support vector machine SVM models are trained using the training sample data set as follows: (4b1) converting the training sample dataset X into vectors, i.e. X = {(xi, yi),..., (x i , y i ), ..., (x N , y N )}, xi is data feature, yi is data label; (4b2) training the model using the training sample dataset X, i.e. X = {(xi, yi),..., (x i , y i )}, xi is data feature, yi is data label; (4b2) call libsvm toolkit, input training sample data X to automatically output the values of optimal parameters gamma and c, and parameters gamma and c represent the performance and fault tolerance of support vector machine SVM respectively; (4b3) After the parameters are set, the optimal values of gamma and c are set, and then the svm-train program in the libsvm tool package is called to obtain a trained binary classification support vector machine SVM model; (4b4) Repeat step (4.2) to obtain all the trained binary classification support vector machine models.

5. The method of claim 1, wherein, In step (5b), the transition relationship among the eight motion states is determined, and a state machine of the vehicle motion state is established to achieve the following (5b1) According to the eight motion states, the following transition relationship among the three motion states is established: Stillness→Vehicle acceleration start→Reverse driving→Vehicle deceleration parking→Stillness; (5b2) According to the transition relationship among the vehicle motion states, a model diagram of the vehicle motion state is drawn.

6. The method of claim 1, wherein, The maximum value δ i (i) of the probability of being in state q t at time t is calculated in step (6) using the Viterbi algorithm, as follows: where δ t-1 (j) is the probability of the state at the previous time step, a ij is the state transition probability from state i to state j, b i (o t ) is the observation probability of state i at time t.

Images