Unmanned vehicle automatic driving scene multi-target tracking method based on online spatial-temporal feature association

Abstract

The invention provides an unmanned vehicle automatic driving scene multi-target tracking method based on online spatial-temporal feature association. The method comprises the steps of data processing, target feature model establishment and spatial-temporal data association. According to the method, the spatial-temporal feature model of the tracking target is constructed, the spatial-temporal feature response of the current frame is detected, online spatial-temporal feature association of the tracking target and the spatial-temporal feature response is achieved, in addition, the optimal association pair between the historical trajectory of the tracking target and the spatial-temporal feature response is solved through the Hungary algorithm, and the tracking accuracy is improved. Therefore, the tracking precision is higher, and the number and motion trails of the tracked targets are directly obtained, so that an important basis is provided for making a decision for an automatic driving system of an unmanned vehicle.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and relates to a multi-object tracking method for autonomous driving scenarios of unmanned vehicles based on online spatio-temporal feature association. Background Technique

[0002] In recent years, with the rapid development of artificial intelligence, many countries have begun to vigorously develop and construct smart cities and intelligent transportation, resulting in a substantial increase in the amount of data in public transportation surveillance videos. Effectively analyzing and utilizing these traffic video data to extract valuable information can assist in optimizing urban traffic and enhancing public safety.

[0003] As a key algorithm of artificial intelligence, the multi-object tracking algorithm is the core in the research field of image and video processing, and has been widely applied in many scenarios such as intelligent transportation, video behavior analysis, aerospace, intelligent robots, and autonomous driving. In the application scenarios of the multi-object tracking algorithm, the trajectories of the tracking objects often change, new tracking objects may enter the scene at any time, some tracking objects may leave the tracking scene, and the similarity between the objects in multi-object tracking is relatively high, and the tracking objects will be occluded during the tracking process. Therefore, the data association between the detection of the tracking objects in the current frame and the existing motion trajectories of the tracking objects in the video surveillance data has become the focus and difficulty of the multi-object tracking algorithm.

[0004] The multi-object tracking algorithm can be divided into a multi-object tracking algorithm based on offline data association and a multi-object tracking algorithm based on online data association according to the data association method. Among them, the multi-object tracking algorithm based on offline data association performs global feature matching according to the spatio-temporal feature responses in all frames. This type of method performs global optimization on the spatio-temporal feature responses of all objects in all frames, and can effectively handle problems such as target loss and identity mis-switching caused by target occlusion and detector misdetection.

[0005] Although the multi-object tracking algorithm based on offline data association can achieve better multi-object tracking performance by using global information, it is difficult to be applied in actual scenarios because only the information of the current and past frames can be used in actual application scenarios, and the information of future frames cannot be obtained. In the process of determining the target trajectory by online data association, the spatio-temporal feature response of the target can be combined with filtering technology, and the Kalman filter technology is used to solve the problems of target position prediction and estimation, and assist in screening out the correct spatio-temporal feature response of the target as the associated object, so as to achieve the purpose of detection and tracking.

[0006] Currently, similarity is calculated based on motion features obtained from visual features and filtering algorithms, and then new tracking targets are associated with existing trajectories in the time dimension. However, the drawback of the above technical solutions is that they do not consider the spatial features of the tracked targets, so the trajectories and quantities of the targets cannot be directly obtained. In addition, spatial information can represent the orientation of the tracking targets in the physical world and defines attributes such as the position and size of the targets. On this basis, combining time information can better understand and predict the behavior of the targets. Therefore, the present invention provides a method for multi-target tracking in an unmanned vehicle autonomous driving scenario based on online spatio-temporal feature association to solve the above technical problems. Summary of the Invention

[0007] The object of the present invention is to provide a method for multi-target tracking in an unmanned vehicle autonomous driving scenario based on online spatio-temporal feature association, which can effectively continuously track multiple targets in time and space, and has high tracking accuracy, providing an important basis for making decisions in unmanned vehicle autonomous driving, and solving the technical problems in the existing autonomous driving tracking technology that the trajectories and quantities of the tracked targets cannot be directly obtained due to the lack of consideration of the spatial features of the tracked targets.

[0008] To this end, the present invention adopts the following technical solutions: A method for multi-target tracking in an unmanned vehicle autonomous driving scenario based on online spatio-temporal feature association, including the following steps:

[0009] a. Perform object detection on the training samples, calibrate the coordinates in the time dimension and the coordinates in the spatial dimension of the detection results, eliminate the detection results with a confidence level less than the set value, and eliminate the overlapping bounding boxes in the detection through the non-maximum suppression algorithm;

[0010] b. After the tracking targets are calibrated in the time and space dimensions, use the Kalman filter to predict the spatial position of each frame of the target;

[0011] c. Obtain the probability of the true spatio-temporal feature response by transforming the spatial position in the target feature model into a search area near the predicted position;

[0012] d. Set the search area as a circle, set the center of the circle as the predicted position of the Kalman filter, and the radius is proportional to the number of frames in which the target is in an unassociated state; specifically:

[0013] Assume that the center position and radius of the search area are (S x , S y ) and R respectively, and the center position coordinates of the rectangular area of the spatio-temporal feature response S s of the tracking target in the current frame are (x, y), then the condition for S s to be located within the search area is:

[0014]

[0015] Among them, R is proportional to N, where N is the number of frames in which the target is in an uncorrelated state;

[0016] e. After determining the search area, extract the spatial and temporal features of the target rectangular area in the training samples, and construct a spatial feature model MS and a temporal feature model MT; where:

[0017] Spatial features refer to obtaining discriminant projection vectors through the recursive linear discriminant analysis algorithm in the tracking target area according to the rectangular image of the tracking target in the training samples. Secondly, the vector is weighted with the color histogram of the area, that is, the spatial features of the target rectangular area in consecutive frames are collected, and the mean value of the above spatial features is used as the target spatial feature model MS, which specifically includes the following steps:

[0018] Establish the initial data set of the recursive linear discriminant analysis algorithm Calculate the initial recursive linear discriminant analysis algorithm model;

[0019] Calculate the initial class mean and overall mean in the tracking target area:

[0020] Calculate the class mean μ for each class c c , calculate the overall mean μ:

[0021]

[0022] Among them, N c is the number of samples in class c; N0 is the total number of samples in the initial data set;

[0023] Calculate the within-class scatter matrix S W :

[0024]

[0025] Among them, T represents transposing the matrix;

[0026] Calculate the between-class scatter matrix S B :

[0027]

[0028] Calculate the initial projection vector: By solving the generalized eigenvalue problem Get the initial projection vector W0;

[0029] For the target area R t Calculate the color histogram H(R t ), and weight and synthesize the projection vector w and the color histogram H(R t ) to obtain the feature vector f t, then the target region R t The spatial feature of is represented as the feature vector f t :

[0030] f t = αw + βH(R t )

[0031] where α and β are both weighting factors, and α + β = 1;

[0032] When the data of the next frame arrives, update the class mean, overall mean, between-class scatter matrix, between-class scatter matrix and projection vector:

[0033] The updated class mean

[0034]

[0035] where x i′ is the feature vector of the i'-th sample belonging to class c in the new dataset; N 1c is the number of samples of class c in the new data;

[0036] The updated overall mean μ new :

[0037]

[0038] where N0 is the total number of samples in the initial dataset; N1 is the total number of samples in the new dataset;

[0039] The updated within-class scatter matrix

[0040]

[0041] where i' ∈ c represents the set of sample indices belonging to class c;

[0042] The updated between-class scatter matrix

[0043]

[0044] By solving the generalized eigenvalue problem Obtain the initial projection vector W new ;

[0045] Obtain the target space feature model MS:

[0046]

[0047] The temporal feature refers to using the histogram of optical flow to describe the magnitude and direction of the motion speed of the tracked object; constructing the temporal feature model MT includes: First, calculate the optical flow feature of the tracked object area through the optical flow algorithm to obtain the optical flow magnitude and direction of each pixel point in this area; Second, equally divide the entire area into R direction intervals between 0° and 360°, and statistically accumulate the optical flow features in each direction interval in the entire area; Finally, use a histogram to represent the optical flow histogram feature model of the object; specifically, it includes the following steps:

[0048] Calculate the optical flow feature vector of each pixel point in the target area R using the optical flow algorithm t in:

[0049] For each pixel point (x, y), calculate the image gradient I in the horizontal direction x , the image gradient I in the vertical direction y and the temporal gradient I t respectively:

[0050]

[0051] Assume that the optical flow feature vector of the pixel point (x, y) at time t is (u, v), then the optical flow equation is:

[0052] I x u + I y v + I t = 0

[0053] Accumulate the optical flow equation within a window to construct a system of linear equations:

[0054]

[0055] where W is the neighborhood window centered on (x, y).

[0056] Express the above system of equations in matrix form:

[0057] Av = b

[0058] where

[0059]

[0060] Solve the system of linear equations to obtain the optical flow feature vector v = (uv);

[0061] For each pixel point (x, y), calculate the magnitude |v(x, y)| and direction θ(x, y) of the optical flow respectively:

[0062]

[0063] θ(x, y) = arctan 2(v(x, y), u(x, y))

[0064] Statistically accumulate the magnitudes of the optical flow features in each direction interval:

[0065] Divide the range [0, 360)° into R direction intervals, each with a size of:

[0066]

[0067] Initialize an array hist of size R θ , for storing the accumulated optical flow magnitudes in each direction interval:

[0068] histθ[i] = 0, 0 ≤ i < R

[0069] Statistically process each pixel point (x, y) in the target region R t to determine the direction interval i it belongs to according to its direction θ(x, y):

[0070]

[0071] Update the accumulated optical flow magnitude in this direction interval:

[0072] hist θ [i] = ∑|v(x, y)|

[0073] Finally, obtain the optical flow histogram feature model, i.e., the temporal feature model MT:

[0074] MT = hist θ

[0075] f. Fuse the spatial feature model MS and the temporal feature model MT of the tracking target rectangular region to obtain the spatio-temporal feature model M of the tracking target;

[0076] For the i-th tracking target T in the j-th frame i , its spatio-temporal feature model is expressed as:

[0077] g. In the data association stage, adopt a spatio-temporal data association framework, including: online detecting the feature response of the tracking target in the current frame and extracting its spatio-temporal features to obtain a spatio-temporal feature response, and performing data association between the spatio-temporal feature response and the target feature model, and calculating the spatial feature similarity between the spatio-temporal feature response and the target feature model; specifically including the following steps:

[0078] The spatio-temporal feature model of the i-th tracking target T in the j-th frame i is:

[0079]

[0080] The spatio-temporal feature response of the tracking target in the current frame is:

[0081]

[0082] where is the spatial feature, is the temporal feature;

[0083] Calculate the spatio-temporal feature response and the similarity S1 between the tracking target T i and the spatial feature model:

[0084]

[0085] Use the Bhattacharyya distance to calculate the similarity S2 between the spatio-temporal feature response and the tracking target T i and the temporal feature model:

[0086]

[0087] where δ is the standard deviation of the Gaussian function, represents the Bhattacharyya distance of the optical flow histogram feature between the spatio-temporal feature response and the tracking target T i ;

[0088] According to the similarities S1 and S2, obtain the similarity metric matrix S:

[0089]

[0090] The elements in the similarity metric matrix are:

[0091]

[0092] h. Optimize the similarity metric matching matrix through the Hungarian algorithm to obtain the optimal spatio-temporal data association result: Specifically, it includes the following steps:

[0093] h1. Subtract the minimum value of each row in the matrix S from that row to obtain a new matrix S':

[0094] S' ij = S ij - min(S i1 , S i2 , …, S in )

[0095] h2. Subtract the minimum value of each row in the matrix S' from that row to obtain the matrix S'':

[0096] S''j = S' ij - min(S' 1j , S' 2j , …, S' nj )

[0097] h3. Find the minimum number of lines in matrix S″ that cover all zero elements. If the number of lines is equal to n, there is a unique zero - element matching that maximizes the total similarity. If the number of lines is less than n, proceed to the next step:

[0098] h4. Find the smallest uncovered element δ and adjust matrix S″, that is, subtract δ from the uncovered elements and add δ to the elements covered twice;

[0099] h5. Repeat steps h3 and h4 until n lines that cover all zero elements are found;

[0100] h6. Select a matching from the final zero - element positions such that each row and each column has only one matching to construct the optimal matching;

[0101] h6. When the optimal spatio - temporal data association result is less than the set threshold, output the best tracking result. When the optimal spatio - temporal data association result is greater than the set threshold, adjust and update the parameters of the target spatio - temporal feature model so that the target feature model can be optimally associated with the spatio - temporal features of the current frame, where:

[0102] Spatial feature The update formula is:

[0103]

[0104] where, represents the spatial feature of target v at time step k; represents the change in spatial feature obtained from the observation of the current frame;

[0105] Temporal feature The update formula is:

[0106]

[0107] where, represents the temporal feature of target v at time step k, represents the change in temporal feature obtained from the observation of the current frame.

[0108] Adjusting and updating the parameters of the target spatio - temporal feature model can reduce the weight of historical features and enhance the influence of current observation features, thereby improving the association accuracy in subsequent frames.

[0109] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0110] 1. The present invention reduces the influence of false detections and repeated detections by eliminating low-confidence detection results and overlapping bounding boxes, improves the accuracy of detection results, and enhances the reliability of the tracking system.

[0111] 2. The present invention uses a Kalman filter for position prediction, reduces the tracking error caused by the rapid movement or sudden change of the target, and improves the continuity and stability of tracking.

[0112] 3. By integrating temporal and spatial feature models, the present invention can accurately capture the changes of the target in the temporal and spatial dimensions, improves the accuracy and robustness of multi-object tracking, improves the tracking accuracy of the algorithm, performs excellently in dealing with complex scenarios such as occlusion and background change, and improves the overall performance of multi-object tracking.

[0113] 4. The present invention detects the spatio-temporal feature response in the current frame in real time, enables the system to quickly respond to the movement and change of the target, reduces the tracking delay, and improves the tracking real-time performance.

[0114] 5. The present invention uses the Hungarian algorithm to perform optimal matching on the target historical trajectory and spatio-temporal feature response, improves the tracking accuracy, directly obtains the number and motion trajectory of the tracking target, and enhances the decision-making basis of the driverless vehicle. BRIEF DESCRIPTION OF THE DRAWINGS

[0115] Figure 1 is a schematic flowchart of the method of this embodiment;

[0116] Figure 2 is a schematic diagram of the target monitored by the driverless vehicle in this embodiment;

[0117] Figure 3 is a schematic diagram of the verification performance indexes of Comparative Example 1, Comparative Example 2 and this embodiment;

[0118] Figure 4 is the verification result of the first frame of the pedestrian video image in the embodiment;

[0119] Figure 5 is the verification result of the third frame of the pedestrian video image in the embodiment;

[0120] Figure 6 is the verification result of the fourth frame of the pedestrian video image in the embodiment;

[0121] Figure 7 is the verification result of the fifth frame of the pedestrian video image in the embodiment;

[0122] Figure 8 is the verification result of the sixth frame of the pedestrian video image in the embodiment;

[0123] Figure 9Verification result of the 8th frame of the pedestrian video image in the embodiment. Detailed implementation manner

[0124] The technical solution of the present invention will be described below in conjunction with the accompanying drawings and the implementation method.

[0125] As Figure 1 shown, this embodiment provides a multi-object tracking method for an autonomous vehicle driving scenario based on online spatio-temporal feature association, including the following steps:

[0126] a. Perform object detection on the training samples, calibrate the coordinates in the time dimension and the coordinates in the space dimension of the detection results, eliminate the detection results with a confidence level less than the set value, and then, eliminate the overlapping bounding boxes in the detection through the non-maximum suppression algorithm;

[0127] b. After calibrating the tracking target in the time and space dimensions, use the Kalman filter to predict the target space position of each frame;

[0128] c. Obtain the probability of the real spatio-temporal feature response by transforming the spatial position in the target feature model into a search area near the predicted position;

[0129] d. As Figure 2 shown, set the search area as a circle, set the center of the circle as the predicted position of the Kalman filter, and the radius is proportional to the number of frames in which the target is in an unassociated state; specifically, set the center position and radius of the search area to be (S x , S y ) and R respectively, and the center position coordinates of the rectangular area of the spatio-temporal feature response S s of the tracking target in the current frame are (x, y), then the condition for S s to be located within the search area is:

[0130]

[0131] And R is proportional to N, where N is the number of frames in which the target is in an unassociated state;

[0132] e. After determining the search area, extract the spatial and temporal features of the target rectangular area in the training samples, and then construct a spatial feature model MS and a temporal feature model MT;

[0133] f. Fuse the spatial feature model NS and the temporal feature model MT of the tracking target rectangular area to obtain the spatio-temporal feature model M of the tracking target. For the i-th tracking target T i in the j-th frame, its spatio-temporal feature model is expressed as:

[0134] g. In the data association stage, a spatio-temporal data association framework is adopted, which includes online detecting the feature responses of the tracked targets in the current frame and extracting their spatio-temporal features. Secondly, the spatio-temporal feature responses are associated with the target feature model, and the spatial feature similarity between the spatio-temporal feature responses and the target feature model is calculated;

[0135] h. Calculate the temporal feature similarity between the spatio-temporal feature responses and the target feature model, and then calculate the fused similarity metric matching matrix.

[0136] The following uses the MOT2015 test platform to test the effectiveness of the multi-target tracking algorithm for the unmanned vehicle autonomous driving scenario based on online spatio-temporal feature association. The simulation hardware environment is Core 3.70GHz, RAM 32GB, MATLAB R2022a, and 64-bit Windows10 SP1. The test content includes the following steps:

[0137] (1) Build a simulation verification environment on the MOT2015 test platform;

[0138] (2) Set the verification performance metrics:

[0139] Multi-target tracking accuracy, which is used to evaluate the average overlap rate between the estimated trajectory and the ground truth trajectory;

[0140] Multi-target tracking precision, which is used to integrate the cumulative accuracy metric indicators of false positives, false negatives, and identity switches;

[0141] Main tracking, which is used to evaluate the integrity of the tracking trajectory and reflects the target re-detection ability;

[0142] False positives, which represent the number of targets misidentified in the detection results;

[0143] False negatives, which represent the number of targets that actually exist but are not detected;

[0144] Identity switches, which represent the number of times the identity label of the same target jumps during the tracking process.

[0145] (3) Calculate the multi-target tracking accuracy MOTA according to the following formula:

[0146]

[0147] where f represents the index number of the video frame; GT f is the number of ground truth target boxes in the f-th frame; FN f is the number of missed detections in the f-th frame, that is, the number of ground truth targets not detected; FP f is the number of false detections in the f-th frame, that is, the number of non-existent targets detected; IDS fis the number of identity switches for the f-th frame, that is, the number of incorrect jumps in the target tracking identity;

[0148] The multi-object tracking accuracy comprehensively reflects the overall performance of the target tracking system by simultaneously quantifying three types of errors: missed detections, false detections, and ID switches; taking the total number of real targets ∑ f GT f as the benchmark, the cumulative error is normalized to a value in the interval [0,1] to intuitively reflect the tracking accuracy. Specifically, MOTA = 1 indicates zero error, and MOTA = 0 indicates complete failure.

[0149] (4) Respectively, take the tracker update method based on high-confidence key samples as comparative algorithm 1, the multi-object tracking method based on trajectory segment confidence as comparative algorithm 2, and conduct a comparative analysis with this embodiment, and use the performance indicators in step (2) as the basis for comparison. The results are as Figure 3 shown;

[0150] (5) Analyze the Figure 3 results:

[0151] Comparative algorithm 1 has the best incorrect identity switches because it significantly reduces the number of target identity jumps by jointly using the target appearance information obtained from the deep network and the target motion information obtained from the Kalman filter;

[0152] Comparative algorithm 2 has the best multi-object tracking accuracy because it can effectively reduce the tracking errors caused by complex situations such as background changes and frequent occlusions by establishing target trajectory segments and trajectory confidence coefficients;

[0153] Since this embodiment improves the tracking accuracy of the algorithm by fusing the time feature model and the space feature model on the basis of the above algorithms, it has the best multi-object tracking accuracy, main tracking, false positives, and false negatives;

[0154] (6) Further verify the effectiveness of this embodiment:

[0155] Select 525 video images from ADL-Rundle-6 in the MOT2015 database, and verify the 1st, 3rd, 4th, 5th, 6th, and 8th frames of the pedestrian video images. The results are as Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Figure 9 shown.

[0156] Analyze Figures 4 to 9It can be seen that this embodiment can continuously track the movement trajectories of Targets 1 to 7 with high accuracy, indicating that the online spatio-temporal correlation algorithm proposed in this embodiment can effectively track multiple targets in terms of time and space.

Claims

1. A multi-object tracking method for autonomous vehicle driving scenarios based on online spatio-temporal feature association, characterized in that, It includes the following steps: Data processing: Perform object detection on the training samples, calibrate the detection results in the time dimension and the spatial dimension, eliminate the detection results with a confidence level less than the set value, eliminate the overlapping bounding boxes in the detection through the non-maximum suppression algorithm, and determine the initial tracking target according to the detection results to provide input for the establishment of the target feature model; Establishment of the target feature model: Predict the position of the tracking target in the current frame through the Kalman filter, and construct the spatio-temporal feature model M. The spatio-temporal feature model M includes the spatial feature model MS and the temporal feature model MT; Spatio-temporal data association: Online detect the feature response of the tracking target in the current frame, extract its spatio-temporal features to obtain the spatio-temporal feature response, perform data association with the spatio-temporal feature model M to obtain the similarity metric matching matrix, and optimize the matching result through the Hungarian algorithm.

2. The multi-object tracking method for the unmanned vehicle autonomous driving scenario based on online spatio-temporal feature association according to claim 1, wherein In the data processing step: The training samples refer to the video frame data obtained from the actual scene; Calibration in the time dimension means aligning the time dimension coordinates of the detection results with the system timestamp; Calibration in the spatial dimension means eliminating the sensor jitter error by translational compensation of the spatial dimension coordinates of the detection results.

3. A multi-object tracking method for an autonomous vehicle driving scenario based on online spatio-temporal feature association according to claim 1, characterized in that The establishment of the target feature model includes the steps: Predict the position of the target in the current frame through the Kalman filter and set a circular search area with its center at the predicted position; Extract the spatial features of the target rectangular area within the search area and construct a spatial feature model Wherein, N represents the time window length; t is the sequence number of the time frame, and t = 1 represents the initial frame or the starting frame of a certain time window; W new is the updated projection vector; H(R t ) is the color histogram, that is, the color distribution of the target region R t at time t; both α and β are weighting factors, and α + β = 1; Extract the time features of the target rectangular area within the search area and construct a time feature model hist θ [i] = ∑|v(x, y)| In the formula, u(x, y) represents the motion component of the pixel point (x, y) in the horizontal direction in the image; v(x, y) represents the motion component of the pixel point (x, y) in the vertical direction in the image; For the i-th tracked target in the j-th frame, based on the spatial feature model of the target rectangular region and the temporal feature model obtain the spatio-temporal feature model of the tracked target 4. A multi-object tracking method for an autonomous vehicle driving scenario based on online spatio-temporal feature association according to claim 1, characterized in that, The spatio-temporal data association includes the following steps: Online detect the tracking target T in the current frame i for its feature response, extract its spatio-temporal features to obtain a spatio-temporal feature response In the formula, is the spatial feature of the tracking target T i ; is the temporal feature of the tracking target T i ; Calculate the spatio-temporal feature response With the tracking target T i The spatial feature model of The similarity S1 between: Calculate the spatio-temporal feature response using the Bhattacharyya distance and the tracking target T i 's temporal feature model The similarity S2 between them is as follows: In the formula, δ represents the standard deviation of the Gaussian function; represents the spatio-temporal feature response and the optical flow histogram feature Bhattacharyya distance i between the tracking target T; According to the spatio-temporal feature response and the spatial feature model of the tracking target T i to obtain the similarity S1 between the spatio-temporal feature response and the temporal feature model of the tracking target T i to obtain the similarity S2, and obtain the similarity metric matrix S : ij :

5. A multi-object tracking method for an autonomous vehicle driving scenario based on online spatio-temporal feature association according to claim 4, characterized in that, The spatio-temporal data association also includes the following steps: Calculate the tracking target T using the Hungarian algorithm i and the spatio-temporal feature response to obtain the best correlation between them, get the optimal spatio-temporal data association result, and obtain the motion trajectory of the tracking target; When the optimal spatio-temporal data association result is less than the set threshold, the best tracking result is output; When the optimal spatio-temporal data association result is greater than the set threshold, the parameters of the target spatio-temporal feature model are adjusted and updated so that the target feature model can be best associated with the spatio-temporal features of the current frame.

6. A multi-object tracking method for an autonomous vehicle driving scenario based on online spatio-temporal feature association according to claim 5, characterized in that The adjustment and update of the parameters of the target spatio-temporal feature model include the following steps: Update the spatial features as follows: In the formula, represents the spatial feature of the target v at time step k; represents the change in the spatial feature obtained from the observation of the current frame; Update the time feature as follows: Among them, represents the temporal feature of target v at time step k; represents the change in temporal feature obtained from the observation of the current frame.

Images