Perception prediction system and prediction method based on point cloud information

Abstract

The perception prediction system and prediction method based on point cloud information belong to the field of information security technology. The present invention solves the problem of low accuracy of the existing method in predicting the trajectory of moving targets. The present invention uses 3D point cloud data as the input of the system, performs real-time detection and recognition of the target object through a convolutional neural network, and dynamically tracks the target based on the recognition result. At the same time, the system of the present invention designs a compensation feedback module to improve the continuity of target tracking. The present invention designs a stabilization module to make the trajectory generated by the system tracking more stable. Based on the tracking trajectory, the present invention uses a Transformer structure to design a prediction module to predict the future trajectory of the identified object. The method of the present invention can be applied to the field of information security technology.

Description

Technical Field

[0001] The present invention belongs to the field of information security technology, and specifically relates to a perception prediction system and prediction method based on point cloud information. Background Art

[0002] Intelligent unmanned systems integrate control, perception, planning, mapping, and artificial intelligence technologies. As a key area of ​​future development, the development of key technologies for robotics and intelligent unmanned systems is receiving increasing attention across various fields. Despite numerous achievements in recent years, widespread application of robotics and intelligent unmanned systems in complex and dynamic environments remains a long way off. One of the most important reasons for this is the lack of effective safety assurance for interactions with the surrounding environment. For example, the perception and avoidance of pedestrians and vehicles by robotic platforms operating in urban environments remains a critical issue that needs to be addressed. Therefore, improving the accuracy of trajectory prediction for moving targets is essential. However, existing methods for trajectory prediction still have low accuracy.

[0003] In the field of robotics and intelligent unmanned systems, the system's perception and prediction of the surrounding environment plays a vital role in path planning and decision-making. With the development of deep learning technology, neural networks have been widely applied to related fields such as point cloud-based object detection and trajectory prediction. Based on these related technologies, designing a point cloud-based perception and prediction system to predict the future trajectory of moving objects has important scientific significance. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem of low accuracy of moving target trajectory prediction in existing methods and to propose a perception prediction system and prediction method based on point cloud information.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] According to one aspect of the present invention, a perception prediction system based on point cloud information includes a target detection module, a tracking module, a compensation feedback module, a stabilization module, and a trajectory prediction module, wherein:

[0007] The target detection module is used to detect and identify moving targets in real time based on the input 3D point cloud data;

[0008] The tracking module is used to track the target based on the detection and recognition results of the target detection module;

[0009] The compensation feedback module is used to store the tracking results, determine whether the newly arrived tracking results and the stored tracking results are from the same tracking object, and feed the determination result back to the tracking module, and then store the newly arrived tracking results;

[0010] The stabilization module is used to process the tracking result of the tracking module, and the trajectory prediction module is used to predict the trajectory of the moving target according to the processing result of the stabilization module.

[0011] According to another aspect of the present invention, a perception prediction method based on point cloud information comprises the following steps:

[0012] The target detection module is used to process the input 3D point cloud data to obtain real-time detection and recognition results of moving targets;

[0013] The tracking module is used to process the real-time detection and recognition results of moving targets to obtain real-time target tracking results;

[0014] The compensation feedback module is used to determine whether the newly arrived tracking result and the stored tracking result are from the same tracking object, and the judgment result is fed back to the tracking module in real time. The tracking module continues to track the target based on the judgment result received in real time. At the same time, the compensation feedback module stores the newly arrived tracking result and the detection result position of the tracking result in the target detection module;

[0015] The stabilization module is used to perform cubic spline interpolation and mean filtering on the tracking results of the tracking module to obtain the processed tracking results, and then the trajectory prediction module is used to predict the trajectory of the moving target based on the processed tracking results.

[0016] The beneficial effects of the present invention are:

[0017] The present invention uses 3D point cloud data as the input to the system, performs real-time detection and identification of target objects through a convolutional neural network, and dynamically tracks the target based on the identification results. Furthermore, the system of the present invention incorporates a compensation feedback module to improve the continuity of target tracking. A stabilization module is also designed to make the trajectory generated by the system more stable. Based on the tracked trajectory, the present invention employs a Transformer structure to design a prediction module to predict the future trajectory of the identified object. Experiments have shown that the system of the present invention can improve the accuracy of moving target trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A block diagram of a perception prediction system based on point cloud information according to the present invention;

[0019] Figure 2 This is the neural network structure diagram of PointPillars and PV-RCNN;

[0020] Figure 3 It is the algorithm flow chart of tracking module;

[0021] Figure 4(a) is the original target tracking trajectory diagram;

[0022] Figure 4(b) shows the target tracking trajectory obtained by spline interpolation;

[0023] Figure 4(c) is the target tracking trajectory after mean filtering;

[0024] Figure 4(d) shows the trajectory after adding Gaussian noise;

[0025] Figure 5 This is the Transformer network structure diagram;

[0026] Figure 6 This is a comparison chart of trajectory tracking results before and after adding the feedback compensation module and without adding the feedback compensation module;

[0027] Figure 7 It is the overall output result diagram of the system of the present invention. DETAILED DESCRIPTION

[0028] Specific implementation method 1: Combination Figure 1 and Figure 3 This embodiment describes a perception prediction system based on point cloud information, which includes a target detection module, a tracking module, a compensation feedback module, a stabilization module, and a trajectory prediction module, wherein:

[0029] The target detection module is used to detect and identify moving targets in real time based on the input 3D point cloud data;

[0030] The tracking module is used to track the target based on the detection and recognition results of the target detection module;

[0031] The compensation feedback module is used to store the tracking results, determine whether the newly arrived tracking results and the stored tracking results are from the same tracking object, and feed the determination result back to the tracking module, and then store the newly arrived tracking results (for subsequent determination);

[0032] The stabilization module is used to process the tracking result of the tracking module, and the trajectory prediction module is used to predict the trajectory of the moving target according to the processing result of the stabilization module.

[0033] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that the target detection module is implemented based on the PointPillars algorithm of the convolutional neural network.

[0034] In order to ensure the real-time performance and accuracy of the system, the present invention adopts the PointPillars algorithm based on convolutional neural network for real-time detection, and adopts the PV-RCNN algorithm with higher accuracy to ensure the training accuracy of the system. The neural network structure of PointPillars and PV-RCNN is as follows: Figure 2 As shown in the figure. PointPillars projects 3D point clouds into "pillars" and uses a two-dimensional CNN to detect targets, thereby reducing the amount of computation and increasing the forward propagation speed. At the same time, based on the increasing maturity of two-dimensional detection technology, relatively good detection accuracy can be achieved even if some features have been lost. Compared with PointPillars, PV-RCNN directly uses a 3D voxel CNN to extract global features based on the 3D point cloud, so it can encode richer contextual information to accurately estimate the credibility and location of the object. In contrast, due to the use of a three-dimensional voxel CNN, PV-RCNN requires more time for forward propagation.

[0035] Other steps and parameters are the same as those in the first embodiment.

[0036] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that the tracking module uses the AB3DMOT algorithm to track the target.

[0037] The present invention uses the AB3DMOT algorithm for target tracking, which combines a 3D Kalman filter and the Hungarian algorithm. It can achieve a tracking speed of 207.4Hz in multi-target tracking tasks. Furthermore, AB3DMOT does not rely on the training results of a neural network.

[0038] Other steps and parameters are the same as those in the first or second embodiment.

[0039] Specific embodiment 4: Based on the prediction method of the perception prediction system based on point cloud information described in specific embodiment 1, the method specifically includes the following steps:

[0040] The target detection module is used to process the input 3D point cloud data to obtain real-time detection and recognition results of moving targets;

[0041] The tracking module is used to process the real-time detection and recognition results of moving targets to obtain real-time target tracking results;

[0042] The compensation feedback module is used to determine whether the newly arrived tracking result and the stored tracking result are from the same tracking object, and the judgment result is fed back to the tracking module in real time. The tracking module continues to track the target based on the judgment result received in real time. At the same time, the compensation feedback module stores the newly arrived tracking result and the detection result position of the tracking result in the target detection module;

[0043] The stabilization module is used to perform cubic spline interpolation and mean filtering on the tracking results of the tracking module to obtain the processed tracking results, and then the trajectory prediction module is used to predict the trajectory of the moving target based on the processed tracking results.

[0044] Specific embodiment 5: This embodiment differs from specific embodiment 4 in that the input vector R of the tracking module is:

[0045] R={frame,x,y,z,rot y ,θ,l,w,h,s,v x ,v y ,v z} (1)

[0046] Among them, frame is the frame number of the detection result, (x, y, z) is the center of the detected target, (l, w, h) is the size of the three-dimensional bounding box where the detected target position is located, and rot y represents the detection result of the target direction, θ represents the heading angle, s is the confidence level, (v x ,v y ,v z ) is the target speed;

[0047] The output vector C of the tracking module is:

[0048] C={frame,id,x,y,z,rot y ,θ,l,w,h,s} (2)

[0049] Among them, id is the tracking result.

[0050] Other steps and parameters are the same as those in the fourth embodiment.

[0051] Specific embodiment 6: This embodiment is different from specific embodiment 4 or 5 in that the heading angle θ is calculated as follows:

[0052]

[0053] Among them, whileθ>π:,θ:=θ-2π means: if the value of θ is greater than π, the operation of θ minus 2π is executed continuously until the condition θ>π is no longer satisfied, and the final value of θ is obtained; else whileθ<-π:,θ:=θ+2π means: if the value of θ is less than -π, the operation of θ plus 2π is executed continuously until the condition θ<-π is no longer satisfied, and the final value of θ is obtained; if the value of θ is -π≤θ≤π, the value of θ does not need to be processed.

[0054] Other steps and parameters are the same as those in the fourth or fifth embodiment.

[0055] Specific embodiment 7: This embodiment is different from any one of specific embodiments 4 to 6 in that the speed (v x ,v y ,v z )for:

[0056]

[0057] Among them, (v x (t),v y (t),v z (t)) is the speed at time t, R x (t) is the position detection result in the x direction at time t, R y (t) is the position detection result in the y direction at time t, R z (t) is the position detection result in the z direction at time t, Δt is the time interval between two detection results, R x (t-Δt) is the position detection result in the x direction at time t-Δt, R y (t-Δt) is the position detection result in the y direction at time t-Δt, R z (t-Δt) is the position detection result in the z direction at time t-Δt.

[0058] In this embodiment, the x direction, the y direction, and the z direction are three coordinate axes of a spatial rectangular coordinate system.

[0059] The other steps and parameters are the same as those in any one of the fourth to sixth embodiments.

[0060] Specific embodiment eight: This embodiment differs from any one of specific embodiments four to seven in that the compensation feedback module is used to determine whether the newly arrived tracking result and the stored tracking result are from the same tracking object. The specific process is as follows:

[0061] If the detection result position corresponding to the newly arrived tracking result in the target detection module and the detection result position corresponding to the stored tracking result meet the conditions: Then the newly arrived tracking result and the stored tracking result come from the same tracking object, otherwise, they do not come from the same tracking object;

[0062] Among them, (pos x ,pos y ,pos z ) represents the current position of the new tracking object, (x i ,y i ,z i ) represents the stored detection result location, ||·||2 is the 2-norm, n represents the total number of stored detection result locations, and k is the threshold.

[0063] The input of the tracking module in the present invention is the detection result of the target detection. However, in actual systems, the instability and discontinuity of the detection results are inevitable. Therefore, the present invention proposes a compensation mechanism to ensure the stability of the tracking results. The compensation mechanism is based on a priori condition that the detected target will not appear randomly in the point cloud map. Therefore, the ID and the location of the detection result are temporarily stored in the storage unit. The system will check the tracking results of it disappearing. When a new tracking object appears, the system will compare the new tracking object with the object stored in the system. Experiments show that this compensation mechanism can well solve the connection problem between the target detection module and the tracking module.

[0064] The other steps and parameters are the same as those in any one of the fourth to seventh embodiments.

[0065] Specific embodiment 9: This embodiment differs from any one of specific embodiments 4 to 8 in that the specific process of the cubic spline interpolation is as follows:

[0066] For the tracking result from a certain tracking target, t is used to represent the discrete time node, the value of the function f at the discrete point is F=f(t), and the domain is divided into {(t0,t1),(t1,t2),...,(t n-1 ,t n′ )}, endpoint t i′ The function value at is f(t i′ )=(x i′ ,y i′ ,z i′ ), i′=0,1,2,3...n′, n′ represents the number of tracking trajectory points of the tracking target. In each sub-domain interval, the interpolation condition S(F) satisfies formula (5). S(F) is a piecewise function containing the interpolation function in each domain:

[0067]

[0068] S'(F), S”(F), S”'(F) satisfy the boundary conditions of formulas (6) to (8):

[0069] S”(F0)=S”(F n′ )=0 (6)

[0070] S0'(F0)=A,S n′-1 '(F n′ )=B (7)

[0071] S″′0(F0)=S″′1(F1),S″′ n′-2 (F n′-1 )=S″′ n′-1 (F n′ ) (8)

[0072] Where A and B are constants, S'(F) is the first derivative of S(F), S″(F) is the second derivative of S(F), and S″′(F) is the third derivative of S(F);

[0073] Polynomial coefficients {a i′ ,b i′ ,c i′ ,d i′ |i′=1,2,...,n′} is calculated using formula (9):

[0074]

[0075] Among them, F i′ =f(t i′ ), h i′ =t i′+1 -t i′ Indicates the step size of the independent variable, m i′ is an intermediate variable;

[0076]

[0077] Formula (10) satisfies: m0=m n′ =0.

[0078] During target detection and tracking, each trajectory is unstable, which means that the detection results of two adjacent frames will produce a lot of jitter. In addition, the original detection result data is sparse, as shown in Figure 4(a). The instability problem will affect the accuracy of trajectory prediction. In order to solve the instability problem, the present invention adopts some numerical analysis methods in the stabilization module. The tracked trajectory can be regarded as a three-dimensional sequence, which uses (x, y, z) to represent the position of the object. In order to avoid the Runge phenomenon during data enhancement, the stabilization module uses cubic spline interpolation to improve the tracking results, thereby improving the prediction performance. The interpolated data set is shown in Figure 4(b).

[0079] The other steps and parameters are the same as those in any one of the fourth to eighth embodiments.

[0080] Specific embodiment 10: This embodiment differs from any one of specific embodiments 4 to 9 in that the specific process of the mean filtering is as follows:

[0081]

[0082] Among them, M i′ Represents the mean filtering result of the i′th tracking point in the tracking result, size represents the kernel size, Represents the interval An integer in the range.

[0083] Mean filtering is performed to ensure the smoothness of the dataset in each dimension. The trajectory curve after mean filtering is shown in Figure 4(c).

[0084] The other steps and parameters are the same as those in the fourth to ninth embodiments.

[0085] The trajectory prediction module uses the Transformer as the prediction network. When the neural network trains the prediction module, adding some noise to the training data can improve the robustness of the Transformer model and avoid overfitting. The noise follows a Gaussian distribution with a mathematical expectation of zero. Based on the principle that noise does not have an impact on the original sample across orders of magnitude, the sample distribution deviation is given by formula (12), and the standard deviation s of the noise distribution is given by formula (13), where k is a constant. In this module, k = 41.28. The training dataset with Gaussian noise is shown in Figure 4(d).

[0086]

[0087]

[0088] Considering the adaptability to nonlinear data and faster prediction speed, Transformer is used as the prediction network. Before trajectory prediction, a priori assumption is made: the trajectory in the future is only determined by the previous states, as shown in formula (14).

[0089] P[S t+1 ..S t+n |S t ...S t-m ]=P[S t+1 ..S t+n |S t ...S t-m-k ] (14)

[0090] In this module, the trajectory training dataset is based on the trajectory tracking results on the KITTI dataset. Transformer can be widely used for sequence prediction. Transformer is a network architecture composed of several encoder and decoder modules. The network uses multi-head attention and contains three basic modules: attention module, feedforward fully connected module, and residual connection. Attention can be calculated using formula (15), where Q represents query, V represents value, and K represents keys.

[0091]

[0092] Transformer is trained with the original parameters, namely 6 layers and 8 attention heads. The position of a pedestrian can be represented by (x, y, z), but in the coordinates of the lidar, the value of the y coordinate is basically unchanged. Therefore, a priori assumption is made: y is a constant for a pedestrian in a short time, so the module only predicts (x, z). The module selects Adam as the optimizer. In addition, dropout and normalization are also applied to avoid overfitting and gradient disappearance problems. The Transformer network structure is as follows Figure 5 shown.

[0093] Experimental results

[0094] Real-time Performance Analysis: In this experiment, the real-time performance of the corresponding modules was measured on a GTX 1050 and a Tesla P100. The relevant experimental data is shown in Table 1. This result shows that even on low-efficiency GPUs, the proposed system can still guarantee real-time performance to a certain extent. Considering that the predicted trajectory is relatively accurate in a short period of time, a system is proposed that predicts the trajectory once every four frames. This can increase the system speed to approximately 20.43Hz (Tesla P100) and 12.96Hz (on a GTX 1050), which can exceed the typical lidar scanning frequency (approximately 10Hz).

[0095] Table 1 Operating speed of each module (ms)

[0096]

[0097] Compensation feedback performance analysis: On the KITTI tracking set, it was found that without the compensation module, the tracking results of the same target would change due to the discontinuity of the detection results. This section verifies the effectiveness of the compensation mechanism through experiments. Figure 6 In the figure, tracking results with and without compensation are compared. The three images on the left show that without the compensation module, the ID of the same pedestrian changes three times, from 3 to 6 to 68. After the compensation mechanism is added, the pedestrian's ID remains unchanged, as shown in the three images on the right. Experiments demonstrate that the compensation module stabilizes detection and tracking results.

[0098] For the predicted trajectory, the system expects to receive longer trajectories in the past time of the object, which is more likely to obtain more accurate prediction results. Therefore, the system of the present invention proposes the concept of ATL, that is, the average tracking length, as shown in formula (16). For the same tracking data set, the fewer the number of tracked targets, the longer the tracking length of each target, and the smaller the ATL, which means that the tracking result is more suitable for trajectory prediction. In this experiment, the tracking test set of KITTI sample numbers 0020 to 0026 was selected to verify the performance of the compensation module. The experimental results are shown in Table 2, where the number No.Frame represents the number of frames, AB3DMOT represents the experiment without the compensation module, and k can refer to The results show that the compensation module can improve the continuity of tracking results and extend the average tracking length.

[0099]

[0100] Table 2 ATL changes before and after adding feedback compensation

[0101]

[0102] Performance Analysis of the Stabilization Module: To measure the performance improvement of the stabilization module, we compared training results under three conditions: original tracking data ( / / Method A); tracking data with cubic spline interpolation and mean filtering (Method B); and tracking data with cubic spline interpolation, mean filtering, and Gaussian noise (Method C). Three evaluation metrics were used: MAD, FAD, and the average of MAD and FAD. The experimental results are shown in Table 3.

[0103] Table 3 Performance of stabilization module

[0104]

[0105] To further verify the stabilization module's ability to improve trajectory prediction performance, we conducted experiments on four public datasets: 1) ETH, 2) Hotel, 3) Zara1, and 4) Zara2. As in the previous experiments, MAD and FAD were used as metrics. The method combining the stabilization module with the Transformer was compared with other methods: linear interpolation (Linear), LSTM-based methods (such as S-GAN individual and S-GAN Social), and the Transformer without a stabilization module. The results show that the stabilization module approach improves datasets collected in relatively complex scenarios, such as the ETH dataset. Furthermore, in relatively simple scenarios, traditional methods (such as linear interpolation) achieve better performance. The experimental results are shown in Table 4.

[0106] Table 4 Performance of the stabilization module under different data sets (Validation)

[0107]

[0108] Overall experimental results: The relevant experiments of the system of the present invention are tested on the KITTI tracking set. The input of the system is the original point cloud and the output is the predicted trajectory. Figure 7 The results are projected from 3D coordinates to 2D coordinates for visualization on the image. These results are also displayed on a bird's-eye view of the point cloud. In the image view, the bounding boxes represent the position results, and the discrete points are the predictions projected to image coordinates. In the bird's-eye view of the point cloud, the discrete points of the trajectory represent the system's predicted trajectory, with different bounding boxes representing different detected objects.

[0109] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A perception prediction system based on point cloud information, characterized in that: The system includes a target detection module, a tracking module, a compensation feedback module, a stabilization module and a trajectory prediction module, wherein: The target detection module is used to detect and identify moving targets in real time based on the input 3D point cloud data; The tracking module is used to track the target based on the detection and recognition results of the target detection module; The compensation feedback module is used to store the tracking results, determine whether the newly arrived tracking results and the stored tracking results are from the same tracking object, and feed the determination result back to the tracking module, and then store the newly arrived tracking results; The specific process of determining whether the newly arrived tracking result and the stored tracking result are from the same tracking object is as follows: If the detection result position corresponding to the newly arrived tracking result in the target detection module and the detection result position corresponding to the stored tracking result meet the conditions: Then the newly arrived tracking result and the stored tracking result come from the same tracking object, otherwise, they do not come from the same tracking object; Among them, (pos x ,pos y ,pos z ) represents the current position of the new tracking object, (x i ,y i ,z i ) represents the stored detection result position, ||·||2 is the 2-norm, n represents the total number of stored detection result positions, and k is the threshold; The stabilization module is used to process the tracking result of the tracking module, and the trajectory prediction module is used to predict the trajectory of the moving target according to the processing result of the stabilization module.

2. The perception prediction system based on point cloud information according to claim 1, characterized in that: The target detection module is implemented based on the PointPillars algorithm of convolutional neural network.

3. The perception prediction system based on point cloud information according to claim 2, characterized in that: The tracking module uses the AB3DMOT algorithm to track the target.

4. The prediction method of the point cloud information-based perception prediction system according to claim 1, characterized in that: The method specifically comprises the following steps: The target detection module is used to process the input 3D point cloud data to obtain real-time detection and recognition results of moving targets; The tracking module is used to process the real-time detection and recognition results of moving targets to obtain real-time target tracking results; The compensation feedback module is used to determine whether the newly arrived tracking result and the stored tracking result are from the same tracking object, and the judgment result is fed back to the tracking module in real time. The tracking module continues to track the target based on the judgment result received in real time. At the same time, the compensation feedback module stores the newly arrived tracking result and the detection result position of the tracking result in the target detection module; The compensation feedback module is used to determine whether the newly arrived tracking result and the stored tracking result are from the same tracking object. The specific process is as follows: If the detection result position corresponding to the newly arrived tracking result in the target detection module and the detection result position corresponding to the stored tracking result meet the conditions: Then the newly arrived tracking result and the stored tracking result come from the same tracking object, otherwise, they do not come from the same tracking object; Among them, (pos x ,pos y ,pos z ) represents the current position of the new tracking object, (x i ,y i ,z i ) represents the stored detection result position, ||·||2 is the 2-norm, n represents the total number of stored detection result positions, and k is the threshold; The stabilization module is used to perform cubic spline interpolation and mean filtering on the tracking results of the tracking module to obtain the processed tracking results, and then the trajectory prediction module is used to predict the trajectory of the moving target based on the processed tracking results.

5. The prediction method of the perception prediction system based on point cloud information according to claim 4, characterized in that: The input vector R of the tracking module is: R={frame,x,y,z,rot y ,θ,l,w,h,s,v x ,v y ,v z } (1) Among them, frame is the frame number of the detection result, (x, y, z) is the center of the detected target, (l, w, h) is the size of the three-dimensional bounding box where the detected target position is located, and rot y represents the detection result of the target direction, θ represents the heading angle, s is the confidence level, (v x ,v y ,v z ) is the target speed; The output vector C of the tracking module is: C={frame,id,x,y,z,rot y ,θ,l,w,h,s} (2) Among them, id is the tracking result.

6. The prediction method of the perception prediction system based on point cloud information according to claim 5, characterized in that: The heading angle θ is calculated as follows: Among them, whileθ>π:,θ:=θ-2π means: if the value of θ is greater than π, the operation of θ minus 2π is executed continuously until the condition θ>π is no longer satisfied, and the final value of θ is obtained; else whileθ<-π:,θ:=θ+2π means: if the value of θ is less than -π, the operation of θ plus 2π is executed continuously until the condition θ<-π is no longer satisfied, and the final value of θ is obtained; if the value of θ is -π≤θ≤π, the value of θ does not need to be processed.

7. The prediction method of the perception prediction system based on point cloud information according to claim 6, characterized in that: The speed (v x ,v y ,v z )for: Among them, (v x (t),v y (t),v z (t)) is the speed at time t, R x (t) is the position detection result in the x direction at time t, R y (t) is the position detection result in the y direction at time t, R z (t) is the position detection result in the z direction at time t, Δt is the time interval between two detection results, R x (t-Δt) is the position detection result in the x direction at time t-Δt, R y (t-Δt) is the position detection result in the y direction at time t-Δt, R z (t-Δt) is the position detection result in the z direction at time t-Δt.

8. The prediction method of the perception prediction system based on point cloud information according to claim 7, characterized in that: The specific process of the cubic spline interpolation is: For the tracking result from a certain tracking target, t is used to represent the discrete time node, the value of the function f at the discrete point is F=f(t), and the domain is divided into {(t0,t1),(t1,t2),...,(t n-1 ,t n′ )}, endpoint t i′ The function value at is f(t i′ ), i′=0,1,2,3...n′, n′ represents the number of tracking trajectory points of the tracking target. In each sub-domain interval, the interpolation condition S(F) satisfies formula (5): S'(F), S”(F), S”'(F) satisfy the boundary conditions of formulas (6) to (8): S”(F0)=S”(F n′ )=0 (6) S0'(F0)=A,S n′-1 '(F n′ )=B (7) S0”'(F0)=S1”'(F1),S' n ” ′-2 (F n′-1 )=S' n ” ′-1 (F n′ ) (8) Where A and B are constants, S'(F) is the first derivative of S(F), S"(F) is the second derivative of S(F), and S"'(F) is the third derivative of S(F); Polynomial coefficients {a i′ ,b i′ ,c i′ ,d i′ |i′=1,2,...,n′} is calculated using formula (9): Among them, F i′ =f(t i′ ), h i′ =t i′+1 -t i′ Indicates the step size of the independent variable, m i′ is an intermediate variable; Formula (10) satisfies: m0=m n′ =0.

9. The prediction method of the perception prediction system based on point cloud information according to claim 8, characterized in that: The specific process of the mean filtering is: Among them, M i′ Represents the mean filtering result of the i′th tracking point in the tracking result, size represents the kernel size, Represents the interval An integer in the range.

Images