A method for extracting traversable areas on the water surface using a highly robust unmanned ship based on lidar

Abstract

The present invention relates to the technical field of laser radar sensors and unmanned boat automatic driving, and more specifically, to a method for extracting traversable areas on the water surface using a highly robust unmanned boat based on laser radar. The present invention uses a laser radar sensor to obtain experimental environment data, performs semantic segmentation on point cloud data through a neural network, and then fuses multiple consecutive frames of riverbank point cloud data. Then, through smoothing filtering, an image corrosion method is used to refine the riverbank point cloud, thereby extracting feature points, which are subsequently sorted as control points. Then, in order to mitigate the rapid changes in the riverbank, the control points are smoothed by an unscented Kalman filter. The smoothed control points are fitted using a B-spline curve, and finally the extracted water surface area is filtered based on a particle filter. The robustness and accuracy of extracting traversable areas on the water surface are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of laser radar sensors and autonomous driving technology for unmanned ships, and more specifically, to a method for extracting traversable areas on a water surface using a highly robust unmanned ship based on laser radar. Background Art

[0002] Chinese patent CN108303988A discloses a target recognition and tracking system for an unmanned boat and its working method, which uses 3D laser radar and vector hydrophone to scan surface and underwater obstacles, the industrial computer analyzes the position of surface obstacles, and the hydrophone processor determines the position of underwater obstacles. According to the acquired obstacle position, the industrial computer adopts an obstacle avoidance algorithm to avoid obstacles, that is, the unmanned boat can avoid obstacles on the surface and underwater, achieving the purpose of being able to operate in waters with complex underwater conditions. This invention mainly aims at avoiding obstacles on the surface and underwater. For scenes of narrow rivers, there will be various shapes of river banks and random noise such as water surface ripples. This invention fails to solve these limitations. LiDAR-based recognition and tracking of surface and underwater targets can only identify and analyze fixed obstacles. Small floating objects such as ripples and leaves on the water surface will be mistaken for obstacles, affecting the extraction accuracy. When extracting the traversable area on the water surface based on LiDAR, using LiDAR on a ship to collect data will cause obvious changes in the collected riverbank due to the bumpy hull. If an unmanned boat is traveling on the water, it will be affected by wind, waves, etc. and will cause bumps, making it highly likely that the radar will not be able to scan the riverbank at a certain moment, causing the riverbank point to disappear, which will have a serious impact on the subsequent algorithm for extracting the traversable area. Summary of the invention

[0003] In order to overcome at least one defect in the above-mentioned prior art, the present invention provides a method for extracting navigable areas on the water surface by a highly robust unmanned boat based on laser radar, which effectively improves the robustness and accuracy of extracting navigable areas on the water surface.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for extracting the traversable area on the water surface by a highly robust unmanned boat based on laser radar, comprising the following steps:

[0005] S1. Use SqueezeSeg’s deep learning network to perform semantic segmentation on the point cloud data collected by LiDAR, and divide the point cloud data into three categories: riverbank, vegetation, and bridge;

[0006] S2. Perform smoothing and filtering operations on the riverbank points extracted in step S1, and then obtain control points by image processing; firstly, fuse the riverbank point cloud data of multiple consecutive frames, and then perform smoothing and filtering, and then refine the riverbank points by image erosion method to obtain the characteristic points of the riverbank;

[0007] S3. Sort the characteristic points of the river bank as control points for B-spline fitting;

[0008] S4. Perform unscented Kalman filter UKF smoothing on the sorted control points;

[0009] S5. Use B-spline to fit the control points and draw a curve of the river bank shape;

[0010] S6. The river surface area can be obtained through step 5, and the area occupied by obstacles in the river surface area is marked and particle filtering is performed.

[0011] The method of extracting the traversable area of ​​the water surface by using a highly robust unmanned boat based on a laser radar uses a laser radar sensor to obtain experimental environment data, performs semantic segmentation on the point cloud data through a neural network, and then fuses the riverbank point cloud data of multiple frames in a row. Then, the riverbank point cloud is refined by smoothing filtering and image corrosion method, thereby extracting feature points, which are subsequently sorted as control points. Then, in order to slow down the rapid changes in the riverbank, the control points are smoothed by an unscented Kalman filter. The smoothed control points are fitted with a B-spline curve, and finally, the extracted water surface area is filtered based on a particle filter.

[0012] In view of the complex and changeable water surface conditions, the present invention proposes a method for extracting the traversable area of ​​the water surface by a highly robust unmanned boat based on laser radar, which can be applied not only to the vast sea level, but also to the water surface extraction of narrow rivers. Based on the deep learning network, the point cloud data obtained by the laser radar is semantically segmented, and the point cloud data is divided into three categories: riverbank, vegetation, and bridge. Since the unmanned boat is traveling on the water surface, it will be affected by wind, waves, etc., which will cause turbulence, which will cause the disappearance of the riverbank points. This problem can be solved by fusing the riverbank point cloud of continuous time frames. Then, by smoothing filtering, the riverbank points can maintain a stable change when the unmanned boat is bumpy. Then, the riverbank points are refined by the image corrosion method to obtain the characteristic points of the riverbank. In order to use B-spline fitting later, the characteristic points of the riverbank are sorted by cosine as the control points of the B-spline. In order to prevent the riverbank curve from changing sharply, the ordered control points are smoothed by the unscented Kalman filter UKF. Then, the control points are depicted by the B-spline value-added method. The above operation can obtain the river surface area, and then mark the area occupied by obstacles in the river surface area. Since the extracted water surface area contains fixed obstacles, ripples, small floating objects and other random noises, particle filtering can be used to filter the water surface noise. Traditional particle filtering is often used for multi-target object tracking. Since noise is discontinuous in time and space, particle filtering can be used to track point clouds. If the point cloud cannot be tracked continuously, it means that it is a noise point cloud, thereby improving the robustness and accuracy of extracting the traversable area of ​​the water surface.

[0013] Furthermore, the step S1 specifically includes: first, it is necessary to change the direct output of the lidar point cloud as input, and then perform semantic segmentation based on the end-to-end pipeline of the convolutional neural network and the reconstructed conditional random field CRF to divide the point cloud data into three categories: river bank, vegetation and bridge.

[0014] Furthermore, the step S2 specifically includes: merging the point cloud of the t frame with the point cloud of the previous two frames to obtain Q, mapping Q to the image coordinate system to obtain the grayscale image Gray_Image, and then performing an erosion operation on Gray_Image to remove the misclassified riverbank point cloud according to the geometric features of the riverbank, and then using a mean filter to smooth, assigning pixels with pixel values ​​less than a threshold M to 0 to reduce the merging error, and finally mapping pixels in the image with pixel values ​​greater than 0 to the point cloud coordinate system to obtain the feature points of the riverbank point cloud.

[0015] Furthermore, the image erosion is to convolve the image with a kernel, a point in the kernel is defined as an anchor point, and then the minimum pixel value of the kernel coverage area is extracted to replace the pixel value of the anchor point position; the erosion has the effect of eliminating small objects in the image and smoothing the boundaries of larger objects; the erosion of the structure element B at the position (x, y) in the A image is defined as placing the origin of B at (x, y), and the minimum value of the area in the image A that overlaps with B is expressed as follows:

[0016]

[0017] The origin of the structural element B is accessed to each pixel in the image A, and the eroded image C can be obtained through the above formula.

[0018] Furthermore, in the step S3, the characteristic points of the river bank extracted in step 2 are divided into four quadrants according to the coordinates and are sorted according to the size of the cosine.

[0019] Furthermore, the step S4 specifically includes:

[0020] S41. Prediction: construct a sigma point set through the optimal value x and covariance p of the previous frame, and map the new sigma point set through the state transfer function; predict the state estimate x and covariance P;

[0021] S42. Observe, construct a sigma point set, map it to a new sigma point set through the observation function, and predict the estimated value z and covariance Pz of the observation;

[0022] S43. Update, calculate the Kalman gain K with the covariance matrix of the same state measurement, and use the weighted average of the Kalman gain to obtain the updated state.

[0023] Furthermore, in the step S5: a series of control points, a series of nodes and a series of coefficients obtained in step S4, each coefficient corresponds to a control point, the calculation of the coefficients must ensure a certain continuity condition, all curve segments are connected together to meet a certain continuity condition, and finally multiple Bezier curves are connected to obtain B-spline.

[0024] Furthermore, the step S6 specifically includes the following steps:

[0025] S61. Initialization phase, the particle set is initialized;

[0026] S62. The particles propagate, and the propagation position is obtained by calculating the displacement difference between the previous two frames to obtain the estimated position of the current frame and adding Gaussian white noise;

[0027] S63. Calculate the weight w of the particle. Assuming that the observation value of the current particle is y, calculate the number of point clouds observed by the particle within the observation range in the current frame. and the number of observation point clouds R(x t-1 ), the particle weight is calculated according to the following formula; if it is noise, the number of point clouds observed by the particle will be very small, and the calculated weight will be small accordingly; the formula is:

[0028]

[0029] S64. Resampling: sorting particles according to their weights, and copying the properties of particles with large weights to particles with small weights;

[0030] S65. Prediction: weighted summing of the weights of the particles is performed to obtain a state estimation value, and the current position of the tracked object is predicted; if the object is not detected for several consecutive frames, it means that the point cloud of the object tracked by the particle is noise;

[0031] S66. Update the particle and update the state of the particle using the latest measurement value.

[0032] The present invention also provides an electronic device, comprising:

[0033] Memory for storing computer programs;

[0034] A processor is used to implement the steps of the above method when executing the computer program.

[0035] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described above are implemented.

[0036] Compared with the prior art, the beneficial effects are:

[0037] 1. The present invention uses a laser radar sensor, which has fewer environmental requirements than other methods that use sensors such as cameras to achieve water surface extraction. Cameras are not only sensitive to lighting conditions, but also have relatively strict environmental requirements;

[0038] 2. The method proposed by the present invention expands the scope of application scenarios. Compared with other water surface applications, most of them can only be applied to relatively ideal water surface environments. The experimental scene of the present invention is an irregular narrow river, which greatly expands the scope of application scenarios;

[0039] 3. The proposed method for extracting traversable water surface areas has high accuracy and robustness. Using laser radar to collect data on the water surface also has some limitations. For example, ripples and leaves on the water surface will reflect the laser radar. The present invention uses particle filtering to filter out ripples and other water surface noise points, greatly improving the accuracy of extracting traversable water surface areas;

[0040] 4. The present invention solves the problem that the laser radar data collected by the unmanned boat will change drastically in several consecutive frames due to the turbulence of the unmanned boat. When the unmanned boat is traveling on the water, there will be different degrees of turbulence, which makes it highly likely that the radar installed on the unmanned boat will not be able to scan the river bank at a certain moment, or the number of river bank point clouds scanned will be significantly different from that of the previous moment, which will cause the traversable area extracted using the river bank points to change dramatically in the continuous time frame, seriously affecting the stability and accuracy of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic flow chart of the method of the present invention.

[0042] Figure 2 It is a neural network model based on SqueezeSeg of the present invention.

[0043] Figure 3 It is a schematic diagram of mean filtering of the present invention.

[0044] Figure 4 It is a schematic diagram of image corrosion of the present invention.

[0045] Figure 5 It is the framework diagram of the unscented Kalman filter of the present invention.

[0046] Figure 6 It is a schematic diagram of the B-spline curve of the present invention.

[0047] Figure 7 It is the pseudo code of the algorithm for extracting the river bank feature points of the present invention. DETAILED DESCRIPTION

[0048] The drawings are only for illustrative purposes and should not be construed as limiting the present invention. To better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the drawings. The positional relationships described in the drawings are only for illustrative purposes and should not be construed as limiting the present invention.

[0049] The steps of extracting the traversable area of ​​the water surface based on the laser radar in this invention are divided into three steps: first, semantic segmentation, then extraction of the water surface area, and finally noise filtering of the water surface area. Figure 1 shown.

[0050] Step 1: Use the neural network based on SqueezeSeg to segment the point cloud data obtained by the lidar. The network structure of SqueezeSeg is a typical encoder-skip connection-decoder structure, which is a lightweight FCN. The neural network model is as follows Figure 2 As shown. First, it is necessary to change the direct output of the LiDAR point cloud as input, and then perform semantic segmentation based on the end-to-end pipeline of the convolutional neural network and the reconstructed conditional random field CRF. In order to be applied to three-dimensional LiDAR point clouds, a CNN is designed to accept the transformed LiDAR point cloud and classify it. Then, the CRF model is further refined by reconstructing it into a recurrent neural network RNN, which reduces the parameter size and computational complexity, and is trained end-to-end with the CNN model.

[0051] (1) Point cloud data

[0052] The point cloud data is labeled and divided into three types of objects: riverbank, vegetation, and bridge. The training data consists of 450 frames of point cloud data, containing a total of 13.5 million point clouds. In our dataset, the point clouds of riverbank, vegetation, and bridge account for 25.2%, 58.3%, and 16.5%, respectively.

[0053] (2) Point cloud conversion

[0054] Due to the sparsity and irregularity of point clouds, general 2D CNN cannot process them directly. It is necessary to convert 3D point cloud data into a CNN-friendly data structure and convert the point cloud into a front view by using spherical projection.

[0055] (3) Neural Network

[0056] The neural network SqueezeSeg is derived from SqueezeNet, such as Figure 2As shown in the figure, the network is a lightweight CNN with 50 times fewer parameters. By referencing SqueezeNet, the layer conv1a is transplanted to fire9 for feature extraction, and then max-pooling is used to downsample the intermediate feature map. Fire9 outputs the downsampled feature map to encode the semantics of the point cloud. The SqueezeSeg network uses a deconvolution module to upsample the feature map in the width dimension to obtain full-resolution label predictions for each point. FireModules and fireDeconvs replace the convolution and deconvolution layers to reduce the number of model parameters and calculations.

[0057] (4) Conditional Random Field (CRF) layer

[0058] Accurate point-by-point label prediction requires not only understanding of the high-level semantics of objects and scenes, but also low-level details. Low-level details are lost in downsampling operations such as max pooling, and the label maps predicted by CNNs usually have blurred boundaries. To improve the accuracy of semantic segmentation, CRF is applied as the last RNN layer to refine the label map. If two points in the cloud are adjacent to each other and have similar intensity measurements, they may belong to the same object and therefore have the same label. Conditional random fields (CRFs) can be used to refine the label map generated by CNN.

[0059] Step 2: In this part, the riverbank point cloud data of multiple consecutive frames are fused, and then smoothed by mean filtering. Then, the riverbank points are refined by image erosion to obtain the characteristic points of the riverbank.

[0060] Mean filtering is the most commonly used method in image processing. From the frequency domain point of view, mean filtering is a low-pass filter that removes high-frequency signals, so it can help eliminate sharp noise in the image and achieve image smoothing, blurring and other functions. The ideal mean filter replaces each pixel in the image with the average value calculated from each pixel and its surrounding pixels. Figure 3 As shown in , the dark gray is the central pixel. The average value of the nine pixels around it and itself is calculated, and the average value is used as the central blue pixel value. If the pixel value of the point does not exceed the set threshold, the point will be deleted.

[0061] Image erosion is used to refine the riverbank point cloud. Image erosion is relative to the pixel value. The principle of erosion is to perform specific logical operations on the area corresponding to the binary image at each pixel position. The operation structure is the corresponding pixel of the output image. The operation effect depends on the size and content of the structural element and the nature of the logical operation. That is, a kernel is used to convolve (scan) the image. A point in the kernel is defined as the anchor point, and then the minimum pixel value (black direction) of the kernel coverage area is extracted to replace the pixel value at the anchor point position. Erosion has the effect of eliminating small objects in the image and smoothing the boundaries of larger objects. The erosion of the structural element B at the (x, y) position in the A image is defined as placing the origin of B at (x, y). The minimum value of the area overlapping with B in the image A is expressed as follows:

[0062]

[0063] The origin of the structural element B accesses each pixel in the image A, and the eroded image C can be obtained by the above formula. Figure 4 As shown, the structural element is a 3x3 pixel block, image A is image x, and the corrosion result of B on A is image v.

[0064] The overall process of this part is described as follows: Pseudo code Figure 7 As shown in the figure, in the algorithm, the point cloud of the t frame is merged with the point cloud of the previous two frames to obtain Q (line 12), Q is mapped to the image coordinate system to obtain the grayscale image Gray_Image, and then the Gray_Image is eroded to remove the misclassified riverbank point cloud according to the geometric features of the riverbank. Subsequently, the mean filter is used for smoothing, and the pixels with pixel values ​​less than the threshold M are assigned to 0 to reduce the merging error. Finally, the pixels with pixel values ​​greater than 0 in the image are mapped to the point cloud coordinate system to obtain the feature points of the riverbank point cloud.

[0065] Step 3: Divide the points of riverbank features extracted in step 2 into four quadrants according to coordinates, sort them according to the size of cosine, and use them as subsequent control points.

[0066] Step 4: Figure 5 As shown, the ordered control points are smoothed by unscented Kalman filtering (UKF).

[0067] UKF unscented Kalman filter is developed on the basis of Kalman filter and transformation. It uses lossless transformation to apply Kalman filter under linear assumption to nonlinear system. First, the conversion model is established, and then the model parameters are fitted according to the model and training data, and then the prediction and update phase of Kalman begins. In the prediction phase, the optimal value x of the previous frame is predicted and sampled, and then mapped to a new point set through the transfer function, and then the estimated value and covariance are predicted based on this point set, and the current observation point set is processed in the same way. Finally, the Kalman gain is calculated based on the observation value and variance of the two, and the optimal value of the current frame is updated according to the gain.

[0068] Step 5: Use B-spline to fit the control points obtained in the above steps.

[0069] B-spline curve refers to a special form of representation in numerical analysis, a sub-discipline of mathematics. It is a linear combination of B-spline base curves. The series of control points, a series of nodes and a series of coefficients obtained in the above steps, each coefficient corresponds to a control point, the calculation of the coefficient must ensure a certain continuity condition, all curve segments are connected together to meet a certain continuity condition, and finally multiple Bezier curves are connected to obtain B-spline.

[0070] like Figure 6 As shown in the figure, there are 8 control points connected by line segments in total, and the B-spline curve is formed by a series of 5 cubic Bezier curves. Generally, the lower the degree, the easier it is for the B-spline curve to approximate the control polyline.

[0071] Step 6: Use particle filtering to filter the noise of the unmanned boat surface area extracted in step 5.

[0072] S61. Initialization phase, the particle set is initialized;

[0073] S62. The particles propagate, and the propagation position is obtained by calculating the displacement difference between the previous two frames to obtain the estimated position of the current frame and adding Gaussian white noise;

[0074] S63. Calculate the weight w of the particle. Assuming that the observation value of the current particle is y, calculate the number of point clouds R (s t (i) ) and the number of observation point clouds R(x t-1 ), the particle weight is calculated according to the following formula; if it is noise, the number of point clouds observed by the particle will be very small, and the calculated weight will be small accordingly; the formula is:

[0075]

[0076] S64. Resampling: sorting particles according to their weights, and copying the properties of particles with large weights to particles with small weights;

[0077] S65. Prediction: weighted summing of the weights of the particles is performed to obtain a state estimation value, and the current position of the tracked object is predicted; if the object is not detected for several consecutive frames, it means that the point cloud of the object tracked by the particle is noise;

[0078] S66. Update the particle and update the state of the particle using the latest measurement value.

[0079] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

[0080] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for extracting traversable areas on a water surface using a highly robust unmanned boat based on laser radar, characterized in that: The following steps are involved: S1. Use SqueezeSeg’s deep learning network to perform semantic segmentation on the point cloud data collected by LiDAR, and divide the point cloud data into three categories: riverbank, vegetation, and bridge; S2. Perform smoothing and filtering operations on the riverbank points extracted in step S1, and then obtain control points by image processing; firstly, fuse the riverbank point cloud data of multiple consecutive frames, and then perform smoothing and filtering, and then refine the riverbank points by image erosion method to obtain the characteristic points of the riverbank; The step S2 specifically includes: merging the point cloud of the t frame with the point clouds of the previous two frames to obtain Q, mapping Q to the image coordinate system to obtain the grayscale image Gray_Image, then performing an erosion operation on Gray_Image, removing the misclassified riverbank point cloud according to the geometric features of the riverbank, and then smoothing by using a mean filter, assigning pixels with pixel values ​​less than a threshold value M to 0 to reduce the merging error, and finally mapping pixels with pixel values ​​greater than 0 in the image to the point cloud coordinate system to obtain feature points of the riverbank point cloud; S3. Sort the characteristic points of the river bank as control points for B-spline fitting; in the step S3, the characteristic points of the river bank extracted in step 2 are divided into four quadrants according to the coordinates and sorted according to the size of the cosine value; S4. Perform unscented Kalman filter UKF smoothing on the sorted control points; S5. Use B-spline to fit the control points and draw a curve of the river bank shape; S6. The river surface area can be obtained through step 5, and the area occupied by obstacles in the river surface area is marked and particle filtering is performed.

2. The method for extracting traversable areas on the water surface by a highly robust unmanned boat based on laser radar according to claim 1 is characterized in that: The step S1 specifically includes: first, it is necessary to change the direct output of the lidar point cloud as input, and then perform semantic segmentation based on the end-to-end pipeline of the convolutional neural network and the reconstructed conditional random field CRF to divide the point cloud data into three categories: river bank, vegetation and bridge.

3. The method for extracting the traversable area on the water surface by a highly robust unmanned boat based on laser radar according to claim 2 is characterized in that: The image erosion is to convolve the image with a kernel, in which a point is defined as an anchor point, and then the minimum pixel value of the kernel coverage area is extracted to replace the pixel value of the anchor point position; erosion has the effect of eliminating small objects in the image and smoothing the boundaries of large objects; the erosion of the structure element B at the position (x, y) in the A image is defined as placing the origin of B at (x, y), and the minimum value of the area in the image A that overlaps with B is expressed as follows: The origin of the structural element B accesses each pixel in the image A, and the eroded image C is obtained through the above formula.

4. The method for extracting traversable areas on the water surface by a highly robust unmanned ship based on laser radar according to claim 1, characterized in that: The step S4 comprises: S41. Prediction: construct a sigma point set through the optimal value x and covariance p of the previous frame, and map the new sigma point set through the state transfer function; predict the state estimate x and covariance P; S42. Observe, construct a sigma point set, map it to a new sigma point set through the observation function, and predict the estimated value z and covariance Pz of the observation; S43. Update, calculate the Kalman gain K according to the covariance matrix measured under the same state, and use the weighted average of the Kalman gain to obtain the updated state.

5. The method for extracting traversable areas on the water surface by a highly robust unmanned ship based on laser radar according to claim 1, characterized in that: In the step S5, a series of control points, a series of nodes and a series of coefficients obtained in step S4 are obtained, each coefficient corresponds to a control point, the calculation of the coefficients must ensure a certain continuity condition, all curve segments are connected together to meet a certain continuity condition, and finally multiple Bezier curves are connected to obtain B-spline.

6. The method for extracting traversable areas on the water surface by a highly robust unmanned ship based on laser radar according to claim 1, characterized in that: The step S6 specifically includes the following steps: S61. Initialization phase, the particle set is initialized; S62. The particles propagate, and the propagation position is obtained by calculating the displacement difference between the previous two frames to obtain the estimated position of the current frame and adding Gaussian white noise; S63. Calculate the weight w of the particle. Assuming that the observation value of the current particle is y, calculate the number of point clouds observed by the particle within the observation range in the current frame. and the number of observation point clouds R(x t-1 ), the particle weight is calculated according to the following formula; if it is noise, the number of point clouds observed by the particle will be very small, and the calculated weight will be small accordingly; the formula is: S64. Resampling: sorting particles according to their weights, and copying the properties of particles with large weights to particles with small weights; S65. Prediction: weighted summing of the weights of the particles is performed to obtain a state estimation value, and the current position of the tracked object is predicted; if the object is not detected for several consecutive frames, it means that the point cloud of the object tracked by the particle is noise; S66. Update the particle and update the state of the particle using the latest measurement value.

7. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the method according to any one of claims 1 to 6 when executing the computer program.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images