A Multi-Vehicle Tracking System and Method Based on Inter-Frame Similarity for Roadside LiDAR

Abstract

A multi-vehicle tracking system and method based on inter-frame similarity for roadside lidar includes: an encoder unit, a decoder unit, a regulator unit, an appearance parameter processing unit, a kinematic parameter processing unit, a point cloud cluster similarity calculation unit, and a global trajectory information optimization unit. This invention achieves accurate tracking of multiple target vehicles within the perception range of roadside lidar (including high-mobility conditions) based on an improved Type I Siamese neural network. By selecting appropriate vehicle static parameters, kinematic parameters, and loss weights, the algorithm can effectively utilize the appearance and kinematic features of the target to achieve tracking under high-mobility driving conditions of roadside vehicles, effectively solving the problem of high-mobility vehicle driving in roadside perception scenarios.

Description

Technical Field

[0001] This invention relates to a technology in the field of autonomous driving, specifically a multi-vehicle tracking system and method for roadside lidar based on the inter-frame similarity of vehicle static parameters and kinematic parameters. Background Technology

[0002] Existing multi-vehicle tracking technologies using vehicle-mounted LiDAR are limited by moving vehicles. Vehicle-mounted LiDARs are installed at low heights and have a small sensing range, only tracking the trajectories of nearby vehicles within this limited sensing range. When nearby vehicles perform highly maneuverable maneuvers, vehicle-mounted LiDARs may lose track of the target. In contrast, roadside LiDARs have a larger sensing range and a fixed viewing angle. When a vehicle enters the sensing range of a roadside LiDAR, regardless of the type of operation the vehicle performs within that range, the vehicle should be located and tracked by the roadside LiDAR. Existing multi-vehicle tracking algorithms based on vehicle-mounted LiDARs struggle to handle the highly maneuverable conditions encountered by roadside LiDARs, resulting in poor multi-vehicle tracking accuracy. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies in handling highly maneuverable vehicles in roadside sensing scenarios. It proposes a multi-vehicle tracking system and method based on inter-frame similarity for roadside lidar. The system utilizes an improved Type I Siamese neural network to accurately track multiple target vehicles within the roadside lidar's sensing range (including in highly maneuverable conditions). By selecting appropriate vehicle static parameters, kinematic parameters, and loss weights, the algorithm effectively leverages the appearance and kinematic features of the targets to achieve tracking in highly maneuverable roadside vehicle scenarios, thus effectively solving the problem of highly maneuverable vehicles in roadside sensing scenarios.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a multi-vehicle tracking system for roadside lidar based on inter-line frame similarity, comprising: an encoder unit, a decoder unit, a modulator unit, an appearance parameter processing unit, a kinematic parameter processing unit, a point cloud cluster similarity calculation unit, and a global trajectory information optimization unit. The encoder unit processes the clustered point cloud information using a three-layer one-dimensional CNN neural network with ReLU and BN layers to obtain a 128-dimensional latent vector. The decoder unit decodes the 128-dimensional latent vector z = Φ(x) from the encoder output into an M×3 value, ultimately reconstructing the shape of M three-dimensional points. The regulator unit uses the chamfer distance as a comparison model shape based on the network's completion loss information. With the shape of the reconstructed model The spacing standard is used to enhance the latent space learned by the Siamese network to preserve meaningful semantic shape information. The appearance parameter processing unit assigns similarity weights based on the center position and length, width, and height information of the vehicle bounding box. The kinematic parameter processing unit assigns similarity weights based on the velocity and orientation angle of the vehicle bounding box. The point cloud cluster similarity calculation unit calculates similarity based on the parameters obtained from the appearance parameter processing unit and the kinematic parameter processing unit with increased weights, obtains the similarity comparison results between each point cloud cluster in the current frame and each point cloud cluster in the next frame, and obtains the association information of the point cloud with the highest similarity. The global trajectory information optimization unit analyzes the interference between point cloud clusters in neighboring frames based on the association results of point cloud clusters in the point cloud cluster similarity calculation unit, sets the result with the highest similarity as the optimal solution, and returns the suboptimal solutions with interference to the point cloud cluster similarity calculation unit for iterative calculation, finally obtaining the global optimal result of multi-vehicle trajectory.

[0006] This invention relates to a multi-vehicle tracking method for roadside lidar based on inter-line frame similarity using the aforementioned system, comprising the following steps:

[0007] Step 1) First, process the point cloud of frame T. Filter the background of the point cloud of frame T using the recorded system background point cloud to obtain the dynamic point cloud. Cluster the dynamic point cloud to obtain vehicle point cloud clusters, and generate 3D bounding boxes based on the point cloud range. When M 3D bounding boxes (with center point, length, width, height, and orientation angle information) are finally labeled in the point cloud of frame T, if there are N vehicles and corresponding N 3D bounding boxes (with center point, length, width, height, orientation angle, and vehicle speed information) in the point cloud of frame T-1, then input each of the M 3D bounding boxes in the point cloud of frame T and each of the N 3D bounding boxes in the point cloud of frame T-1 into the tracking system for similarity calculation.

[0008] Step 2) The tracking system feeds the two inputs into a CNN network for calculation, and finally obtains the similarity between the two inputs under set parameters (from 0 to 1). This invention improves upon this, so that it can be applied to laser point clouds and can perform comprehensive calculations on vehicle static parameters (bounding box center point position, length, width, height, and orientation angle) and kinematic parameters (vehicle speed).

[0009] This invention adds different loss weights ki to each parameter. The M 3D bounding boxes in the T-th frame point cloud are respectively processed by the tracking system with the N 3D bounding boxes in the T-1-th frame point cloud (each of the M 3D bounding boxes in the T-th frame point cloud calculates the straight-line distance with the N 3D bounding boxes in the T-1-th frame point cloud to obtain the calculated vehicle speed). Each of the N 3D bounding boxes in the T-1-th frame point cloud can obtain a 3D bounding box in the T-th frame point cloud with the highest similarity, and the connection between the two is regarded as a temporary trajectory.

[0010] Step 3) After obtaining N temporary trajectories between the point cloud of frame T-1 and the point cloud of frame T, compare the similarity of the trajectories where there is interference at the connection position of the N trajectories. Take the trajectory with higher similarity as the real trajectory, and take the second-best similarity value associated in the previous step as the temporary trajectory for the corresponding false trajectory. Perform the above comparison again, and finally obtain L real trajectories with the highest similarity and record them.

[0011] When a 3D bounding box exists in the point cloud of frame T-1 where no trajectory has been acquired, two cases are discussed: If the historical trajectory line reaches the edge of the roadside perception range in the next frame, it is recorded as a lost vehicle; if it does not reach the edge of the roadside perception range, the predicted trajectory for the next frame is recorded according to the orientation angle and vehicle speed of the 3D bounding box in the previous frame. When a 3D bounding box exists in the point cloud of frame T where no trajectory has been acquired, two cases are discussed: If the vehicle's 3D bounding box is located at the edge of the roadside perception range, it is set as a new vehicle ID; if it is not located at the edge of the roadside perception range, it is recorded as a temporary ID, and an attempt is made to associate it in the point cloud. If the association fails, no vehicle ID is assigned. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the system of the present invention;

[0013] Figure 2 This is a schematic diagram of the point cloud cluster similarity calculation unit and the global trajectory information optimization unit.

[0014] Figure 3 Schematic diagram of encoder unit and decoder unit;

[0015] Figure 4 Schematic diagram of appearance parameter processing unit and kinematic parameter processing unit;

[0016] Figure 5 This is a schematic diagram of the area where the algorithm uses roadside datasets for sample collection. Detailed Implementation

[0017] like Figure 1 As shown in the figure, this embodiment relates to a roadside lidar multi-vehicle tracking system based on inter-frame similarity calculation, including: an encoder unit, a decoder unit, an adjuster unit, an appearance parameter processing unit, a kinematic parameter processing unit, a point cloud cluster similarity calculation unit, and a global trajectory information optimization unit, wherein the information flow direction is from bottom to top, and the processed upper-layer information is applied to multi-vehicle tracking.

[0018] The first layer from bottom to top in the diagram is the point cloud information processing layer. This part belongs to the point cloud cluster similarity calculation unit of the overall structure. The input information of this layer is the raw point cloud information obtained by the roadside lidar in the neighboring frame. After background filtering and clustering, the point cloud clusters of suspected vehicles are obtained (since the perception area of ​​the roadside lidar is fixed, the background can be easily filtered out, and the clustering of vehicle point clouds on the road surface is also a relatively mature method, with many public results available online, so the specific method will not be described here).

[0019] The tracking system calculates the similarity between point cloud clusters and point cloud clusters in adjacent frames. That is, each point cloud cluster in the current frame will be compared with all point cloud clusters in the neighboring frames, and the result with the highest similarity will be recorded and associated.

[0020] The second and third layers from bottom to top belong to the global trajectory information optimization unit of the overall structure. The second layer from bottom to top is the trajectory information comparison layer. The association result of the maximum similarity between the current frame point cloud cluster and the neighboring frame point cloud cluster generated by the first layer structure is recorded as the candidate trajectory. At this time, multiple current frame point cloud clusters are allowed to be associated with the same point cloud clusters in neighboring frames.

[0021] The third layer from bottom to top is the topology verification layer. Road topology occupancy analysis will be performed between the candidate trajectories. If there is a conflict in the trajectory, the trajectory with the highest similarity will be recorded as the real trajectory, while the other trajectory that caused the conflict will return to the first layer for iterative calculation. The second highest similarity association between the point cloud clusters will be selected as the trajectory, and the above steps will be repeated.

[0022] like Figure 2 As shown, the encoder unit processes the clustered point cloud information using a three-layer one-dimensional CNN neural network with ReLU and BN layers to obtain a 128-dimensional latent vector. The decoder unit decodes the 128-dimensional latent vector z = Φ(x) from the encoder output into an M×3 value, ultimately reconstructing the shape of M three-dimensional points. The regulator unit uses the chamfer distance as a comparison model shape based on the network's completion loss information. With the shape of the reconstructed model The spacing criterion is used to enhance the latent space learned by the Siamese network to preserve meaningful semantic shape information.

[0023] like Figure 3 As shown, the encoder unit includes: a one-dimensional convolutional layer, a ReLU layer with filter size [64, 128, K], and a BN layer. This unit takes a series of point cloud clusters as input, where the point cloud data contains a given object and an initial 3D bounding box corresponding to the object's position in the first frame. For a given frame t, a set of candidate shapes... Encoding to latent vector In the middle, then with the shape from the model latent vectors The best candidate object is selected as the object in the current frame and the model shape x0 is updated accordingly.

[0024] To obtain K-dimensional latent vectors, K=128 is a suitable latent vector dimension because it optimally balances computational efficiency, latent space compactness, and tracking performance. The network's input is preprocessed to 2048 points by randomly dropping or copying points; therefore, mini-batch training is chosen during the training process.

[0025] This embodiment uses KITTI as a commonly used dataset, where over 96% of vehicle point clouds contain fewer than 2048 points. Therefore, the aforementioned tracking system can be well applied to most point cloud datasets. This section includes approximately 25K parameters. The encoder unit takes a point cloud of N = 2048 points as input. A three-layer one-dimensional CNN neural network with ReLU and BN layers encodes the point cloud into a 128-dimensional latent vector.

[0026] After the encoder unit extracts the latent meaning vector z from the point cloud data x, it compares a pair of shapes x and... The similarity, latent vector z and Cosine similarity between Where ||.||2 is the L2-norm of the vector.

[0027] When training an encoder unit using any given frame as samples, let x be the point cloud of the target being tracked. This is the ground truth model point cloud obtained by connecting the target point cloud to all frames within the trajectory segment. Based on the loss function, the above tracking system is trained so that the candidate shape x matches the model shape. Regression between the distance function, where: x and The pose is determined by three degrees of freedom in the plane (t) x ,t y The distance difference d(.,.) is determined by the L2 norm ||.||2 of the difference between parameterized poses, where α is in degrees, and is obtained by assigning it a weight of 0.2 to achieve the same result as t in meters. x and t y Approximate scale.

[0028] This embodiment uses the MSE loss function to maximize the similarity between partial shapes and the similarity between the complete shape. The MSE loss function, or Mean Squared Error Loss Function, is a classic loss function that effectively solves regression problems. Specifically, customizing the parameters of the loss function will improve its performance. Wherein: a Gaussian function with parameters μ=0 and σ=1 is selected, where the two parameters represent the mean and variance in the Gaussian function as a differentiable function ρ(), used to weaken the distance between positive and negative samples: when the distance between positive and negative samples is 0, ρ() takes the value of 1, and then as the distance between positive and negative samples increases, ρ() decreases, and finally when the distance between positive and negative samples is infinitely large, ρ() approaches 0.

[0029] The decoder unit comprises two fully connected layers that decode the 128-dimensional latent vector z = ψ(x) into M*3 values, i.e., reconstruct the shape of M three-dimensional points. In this embodiment, the M value of the decoder unit is 2048, the hidden layer size used is 1024, and it includes a total of approximately 6.4M parameters.

[0030] The regulator unit adjusts the changes in model shape during encoding and decoding, using the chamfer distance as a comparison of model shapes. and reconstruct model shape The standard for distinguishing between them is the chamfer distance: in: and These represent the shapes of the reconstructed point cloud model, with the first item representing... From any point in the middle to The sum of the minimum distances, the second term represents From any point in the middle to The sum of minimum distances. The tracking loss function forces the partial shape of the encoded model to approximate the shape of its respective encoded model, while the completion loss forces the encoded model to store semantic information for decoding. Therefore, this regularization is used to strengthen the latent space learned by the I-Siamese network to preserve meaningful semantic shape information.

[0031] like Figure 4As shown, the bounding box center position, bounding box shape, inter-frame vehicle speed, and direction angle of the appearance parameter processing unit and kinematic parameter processing unit are directly obtained from the perception results in the point cloud data and then input into the point cloud cluster similarity calculation unit for calculation. Speed ​​is obtained through two frames of data: given the bounding box information of frame t and frame t+1, the speed of the target object in different directions between the two frames can be calculated based on the offset between their bounding box positions and the time interval between the two frames. This speed is then used as a kinematic parameter and input into the point cloud cluster similarity calculation unit for calculation. Finally, the improved parameters based on the original similarity matrix are (x, y, w, h, l, v). x ,v y ,α), where: x, y represent the center position of the bounding box, w, h, l represent the shape of the bounding box, v x ,v y α represents vehicle speed, and α represents the direction angle. A point cloud cluster similarity calculation unit is introduced. Figure 1 The middle and upper-level structures perform further calculations.

[0032] To obtain the target vehicle's speed information, this embodiment assumes a constant-speed motion model for the vehicle between frames. For adjacent given frames t-1 and t, the target object's velocity and position satisfy the following relationship: (x, y) t =(x,y) t-1 +(v x ,v y ) t-1 ·Δt+∈ (x,y) , Where: the process noise of each state variable ∈ (x,y) and It follows a normal distribution with a mean of 0. Considering that the size of the target object can introduce errors in position calculation, the variance of the position noise is... The variance of velocity noise changes with the size of the target object. It is inversely proportional to the number of frames successfully tracked.

[0033] In this embodiment, Δt is set to 0.1 seconds.

[0034] Through specific practical experiments, taking intersections as an example, vehicles may exhibit sudden acceleration or deceleration maneuvers significantly different from those of adjacent vehicles, or make turns with angles exceeding 90 degrees. Under such highly maneuverable conditions, traditional multi-vehicle tracking methods struggle to guarantee accurate vehicle tracking. This invention proposes a novel multi-vehicle tracking method that comprehensively considers vehicle kinematic and shape parameters, effectively solving the accuracy problem of multi-vehicle tracking under highly maneuverable conditions in roadside lidar sensing.

[0035] In a specific embodiment, the algorithm of this invention was tested on the Tsinghua University public dataset DAIR-V2X and self-collected data segments. The test results show that the method proposed in this invention exhibits good performance in both comprehensive operating conditions and high-mobility operating conditions.

[0036]

[0037]

[0038] Compared to existing methods, this method demonstrates excellent tracking performance under both comprehensive and high-maneuverability conditions. Its tracking accuracy (MOTA) and tracking precision (MOTP) are superior to existing methods, especially under high-maneuverability conditions, where this method shows a significant advantage over existing methods.

[0039]

[0040]

[0041] The SAIKF operating conditions in the table above are recorded by: Khalkhali, Maryam Baradaran; Vahedian, Abedin; Yazdi, Hadi Sadoghi, Situation Assessment-Augmented Interactive Kalman Filter for Multi-Vehicle Tracking. IEEE transactions on intelligent transportation systems, pp. 1-11, 05 Feb 2021.

[0042] PCGDA 3D Multi-Object Tracking working conditions are recorded in: Wu, Hai, Wenkai Han, ChengluWen, Xin Li, and Cheng Wang. "3D Multi-Object Tracking in Point Clouds Based on Prediction Confidence-Guided Data Association." IEEE Transactions on Intelligent Transportation Systems (2021): 1-10.

[0043] This invention comprehensively utilizes vehicle static parameters and kinematic parameters to calculate inter-frame similarity in order to complete multi-vehicle tracking based on roadside lidar. The tracking accuracy of this method in roadside lidar perception scenarios is significantly better than existing methods.

[0044] This invention achieves accurate tracking of multiple target vehicles within the perception range of a roadside lidar by improving the K-Siamese network. This network can effectively ensure tracking accuracy even in scenarios where vehicles are highly maneuvering within the roadside perception range.

[0045] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A multi-vehicle tracking system for roadside lidar based on inter-line frame similarity, characterized in that, include: The system comprises an encoder unit, a decoder unit, a regulator unit, an appearance parameter processing unit, a kinematic parameter processing unit, a point cloud cluster similarity calculation unit, and a global trajectory information optimization unit. Specifically: the encoder unit processes the clustered point cloud information using a three-layer one-dimensional CNN neural network with ReLU and BN layers to obtain a 128-dimensional latent vector; the decoder unit then processes the 128-dimensional latent vector output by the encoder. Decoded as The final reconstruction includes the shapes of M three-dimensional points. The regulator unit uses the chamfer distance as a comparison model shape based on the network's completion loss information. With the shape of the reconstructed model The spacing standard is used to enhance the latent space learned by the Siamese network to preserve meaningful semantic shape information; the appearance parameter processing unit assigns similarity weights based on the center position and length, width, and height information of the vehicle bounding box; the kinematic parameter processing unit assigns similarity weights based on the velocity and orientation angle of the vehicle bounding box; the point cloud cluster similarity calculation unit calculates similarity based on the parameters obtained from the appearance parameter processing unit and the kinematic parameter processing unit with increased weights, obtains the similarity comparison results between each point cloud cluster in the current frame and each point cloud cluster in the next frame, and obtains the point cloud association information with the highest similarity; the global trajectory information optimization unit analyzes the interference between point cloud clusters based on the association results of neighboring frame point cloud clusters obtained from the point cloud cluster similarity calculation unit, sets the result with the highest similarity as the optimal solution, and returns the suboptimal solutions with interference to the point cloud cluster similarity calculation unit for iterative calculation, finally obtaining the global optimal result of multi-vehicle trajectory; The global trajectory information optimization unit includes: The trajectory information comparison layer records the association results of the maximum similarity between the current frame point cloud cluster and the neighboring frame point cloud cluster generated by the point cloud cluster similarity calculation unit as the candidate trajectory. At this time, multiple current frame point cloud clusters are allowed to be associated with the same point cloud clusters in neighboring frames. In the topology verification layer, road topology occupancy analysis will be performed between candidate trajectories. If there is a conflict between the trajectories, the trajectory with the highest similarity will be recorded as the real trajectory, while the other trajectory that caused the conflict will be returned to the first layer for iterative calculation. The association result with the second highest similarity between point cloud clusters will be reselected as the candidate trajectory, and the above steps will be repeated. The regulator unit adjusts the changes in model shape during encoding and decoding, using the chamfer distance as a comparison of model shapes. and reconstruct model shape The standard for distinguishing between them is the chamfer distance: ,in: and These represent the shapes of the reconstructed point cloud model, with the first item representing... From any point in the middle to The sum of the minimum distances, the second term represents From any point in the middle to The minimum distance sum; the tracking loss function forces the partial shape of the encoded model to be similar to the shape of each of the encoded models, and the completion loss forces the encoded model to store semantic information for decoding.

2. The multi-vehicle tracking system for roadside lidar based on inter-line frame similarity as described in claim 1, characterized in that, The encoder unit includes a one-dimensional convolutional layer, a ReLU layer with filter sizes of [64, 128, K], and a BN layer. This encoder unit takes a series of point cloud clusters as input, where the point cloud data contains a given object and an initial 3D bounding box corresponding to the object's position in the first frame. For a given frame t, a set of candidate shapes... Encoding to latent vector In the middle, then with the shape from the model latent vectors The model is compared, the best candidate object is selected as the object in the current frame, and the model shape is updated accordingly. .

3. The multi-vehicle tracking system for roadside lidar based on inter-line frame similarity according to claim 1 or 2, characterized in that, After extracting the latent meaning vector z from the point cloud data x, the encoder unit compares a pair of shapes x and z. Similarity, latent vector z and Cosine similarity between ,in: Let L2 be the L2 norm of the vector. When training an encoder unit using any given frame as samples, let x be the point cloud of the target being tracked. The target point cloud is obtained by connecting it to the real model point cloud within all frames of the trajectory segment; based on the loss function, the above tracking system is trained so that the candidate shape x matches the model shape. Regression between the distance function, where: x and The pose is determined by three degrees of freedom in the plane ( The distance difference d(.,.) is determined by the L2 norm of the difference between parameterized poses. , The unit is angle, and it is weighted by 0.2 to obtain the same value as the unit is meter. and The similarity scale is approximated; the MSE loss function is used to maximize the similarity between partial shapes and the similarity between complete shapes, specifically: Among them: the selected parameter is , The Gaussian function, where: the two parameters represent the mean and variance of the Gaussian function as differentiable functions. This is used to weaken the distance value between positive and negative samples: when the distance between positive and negative samples is 0. The value starts at 1, and then increases with the distance between positive and negative samples. Decreasing, eventually reaching a point where the distance between positive and negative samples becomes infinitely large. Approaching 0.

4. The multi-vehicle tracking system for roadside lidar based on inter-line frame similarity as described in claim 1, characterized in that, The decoder unit includes two fully connected layers that process the 128-dimensional latent vector. Decoding yields an M*3 value, which reconstructs the shape of M three-dimensional points. .

5. A method for multi-vehicle tracking using roadside lidar based on inter-frame similarity of the system described in any one of claims 1-4, characterized in that, Includes the following steps: Step 1): First, process the point cloud of frame T. Filter the background of the point cloud of frame T using the recorded system background point cloud to obtain the dynamic point cloud. Cluster the dynamic point cloud to obtain vehicle point cloud clusters and provide 3D bounding boxes based on the point cloud range. Finally, label M 3D bounding boxes in the point cloud of frame T. When there are N vehicles and corresponding N 3D bounding boxes in the point cloud of frame T-1, input each bounding box in the M 3D bounding boxes of the point cloud of frame T and each bounding box in the N 3D bounding boxes of the point cloud of frame T-1 into the tracking system for similarity calculation. Step 2) The tracking system feeds the two inputs into a CNN network for calculation, and finally obtains the similarity between the two inputs under the set parameters, which is then applied to the laser point cloud and used to perform comprehensive calculations on the vehicle's static parameters and kinematic parameters. Step 3) After obtaining N temporary trajectories between the point cloud of frame T-1 and the point cloud of frame T, compare the similarity of the trajectories where there is interference at the connection position of the N trajectories. Take the trajectory with higher similarity as the real trajectory, and take the second-best similarity value associated in the previous step as the temporary trajectory for the corresponding false trajectory. Perform the above comparison again, and finally obtain L real trajectories with the highest similarity and record them.

6. The multi-vehicle tracking method for roadside lidar based on inter-line frame similarity according to claim 5, characterized in that, Different loss weights ki are added to each parameter; the M 3D bounding boxes in the T-th frame point cloud are respectively processed by the tracking system with the N 3D bounding boxes in the T-1-th frame point cloud, that is, each of the M 3D bounding boxes in the T-th frame point cloud calculates the straight-line distance with the N 3D bounding boxes in the T-1-th frame point cloud to obtain the calculated vehicle speed. Each of the N 3D bounding boxes in the T-1-th frame point cloud can obtain a 3D bounding box in the T-th frame point cloud with the highest similarity, and the connection between the two is regarded as a temporary trajectory.

7. The multi-vehicle tracking method for roadside lidar based on inter-line frame similarity according to claim 5, characterized in that, when When a 3D bounding box exists in the point cloud of frame T-1 where no trajectory has been acquired, two cases are discussed: if the line connecting the historical trajectory in the next frame reaches the edge of the roadside perception range, it is recorded as a lost vehicle; if it does not reach the edge of the roadside perception range, the predicted trajectory for the next frame is recorded according to the orientation angle and vehicle speed of the 3D bounding box in the previous frame. When a 3D bounding box exists in the point cloud of frame T where no trajectory has been acquired, two cases are discussed: if the vehicle's 3D bounding box is located at the edge of the roadside perception range, it is set as a new vehicle ID; if it is not located at the edge of the roadside perception range, it is recorded as a temporary ID, and an attempt is made to associate it in the point cloud. If the association fails, no vehicle ID is assigned.

Images