A novel fast pedestrian recognition algorithm based on point cloud compression and boundary extraction

Point cloud compression method based on boundary extraction

The network structure of the YOLOv5

YOLOv5 network structure comprises input, backbone, neck, and prediction components. The main functions of each component are described as follows:

• (1) Input: YOLOv5 uses the Mosaic data enhancement method in YOLOv4 to enrich data sets, reduce hardware requirements and reduce the use of GPU. The adaptive anchor frame calculation function is embedded in the whole training code. The switch can be adjusted automatically, and the adaptive image scaling can improve the speed of target detection in the reasoning stage.

• (2) Backbone: The focus module is the unique structure of the YOLOv5. The key is to slice the feature map into blocks and then transfer the image features to the next block through multi-layer convolution pooling, cross-stage partial network (CSPNet) (Chien-Yao et al., 2020), and feature pyramid pooling (SPP) (He et al., 2015) structures.

• (3) Neck: The feature pyramid networks (FPN) (Tsung-Yi et al., 2017) and path aggregation network (PAN) (Shu et al., 2018) structures of the YOLOv4 are used.

• However, the CSP2 structure designed concerning CSPnet is adopted, which strengthens information dissemination and can accurately retain spatial information as the neck structure of the YOLOv5 differs.

• (4) Prediction: The improved loss function called GIOU-Loss can accurately identify some objects with overlapping occlusion and resolve the defects of the conventional IOU.

The loss function of the YOLOv5 network

In the YOLOv5 network, $\mathrm{C}\mathrm{I}\mathrm{o}\mathrm{U}$ $\mathrm{C}\mathrm{I}\mathrm{o}\mathrm{U}$ loss in the $\mathrm{I}\mathrm{o}\mathrm{U}$ $\mathrm{I}\mathrm{o}\mathrm{U}$ series loss function is used to calculate the boundary box loss (Box) of the target box and prediction box. The standard and prediction frames’ confidence loss and classification loss are calculated by using the binary cross entropy loss (BCEL) function defined by

• (1) Bounding box loss:

(11) $$CIoU=IoU-\frac{{\rho }^2}{c^2}-\alpha \nu$$

(12) $$\nu =\frac{4}{{\pi }^2}(\arctan \frac{{\omega }^{gt}}{h^{gt}}-\arctan \frac{{\omega }^p}{h^p}{)}^2$$

(13) $$los{s}_{CIoU}=1-CIoU$$where the $\mathrm{I}\mathrm{o}\mathrm{U}$ $\mathrm{I}\mathrm{o}\mathrm{U}$ denotes the ratio of the overlapping area of the prediction box and the target box to the area of the parallel part; $\rho$ denotes the distance between the centre point of the prediction box and the target box; $\nu$ is used to indicate the similarity of aspect ratio; $\alpha$ represents the influence coefficient of $\nu$.

• (2) Confidence loss and classification loss:

(14) $$Loss=-{\sum }_{i=1}^N\left(y_i\log (p\left(x_i\right)+(1-y_i)\log \left(1-p\left(x_i\right)\right)\right)$$where y denotes the real label and $p(x)$ represents the model output.

Mapping of point cloud data to images

The acquisition platform of the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) dataset is equipped with a camera, 64-line lidar, and optical lens (Zhiyong, Luo & Dai, 2017). Mapping the PCD to the rotation matrix of the image requires the calibration parameters of the lidar and camera, the internal parameters of the camera, and the rotation matrix. According to the conversion formula calibrated during the acquisition of the KITTI data set, the correspondence is computed between $X(x,y,z,1)$ the point cloud coordinate and $Y(u,v,1)$ image coordinate system, then, one-to-one correspondence is established between the PCD and the image pixel. Equation (15) expresses this relationship.

(15) $$Y=P_{rect}^{\left(i\right)}{R}_{rect}^{\left(0\right)}{T}_{velo}^{cam}X$$where ${\mathrm{P}}_{\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}}^{\left(\mathrm{i}\right)}$ ${\mathrm{P}}_{\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}}^{\left(\mathrm{i}\right)}$ denotes the internal parameter matrix representing the i^th camera, ${\mathrm{R}}_{\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}}^{\left(0\right)}$ ${\mathrm{R}}_{\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t}}^{\left(0\right)}$ represents the correction matrix from each camera to camera 0, ${\mathrm{T}}_{\mathrm{v}\mathrm{e}\mathrm{l}\mathrm{o}}^{\mathrm{c}\mathrm{a}\mathrm{m}}$ ${\mathrm{T}}_{\mathrm{v}\mathrm{e}\mathrm{l}\mathrm{o}}^{\mathrm{c}\mathrm{a}\mathrm{m}}$ denotes the rotation and translation matrices from the camera to the laser radar. Mapping image pixels to PCD requires each matrix’s row and column expansion. X represents the homogeneous coordinate form of PCD, and ${\mathrm{T}}_{\mathrm{v}\mathrm{e}\mathrm{l}\mathrm{o}}^{\mathrm{c}\mathrm{a}\mathrm{m}}$ ${\mathrm{T}}_{\mathrm{v}\mathrm{e}\mathrm{l}\mathrm{o}}^{\mathrm{c}\mathrm{a}\mathrm{m}}$ denotes the external parameter matrix of laser radar and the camera obtained through calibration, including the rotation matrix and translation matrix.

(16) $$\alpha .\left(\begin{array}{c}u\\ v\\ 1\end{array}\right)=\left(\begin{array}{c}{P}_{00}\\ P_{10}\\ \begin{array}{c}{P}_{20}\\ P_{30}\end{array}\end{array}\begin{array}{c}{P}_{01}\\ P_{11}\\ \begin{array}{c}{P}_{21}\\ P_{31}\end{array}\end{array}\begin{array}{c}{P}_{02}\\ P_{12}\\ \begin{array}{c}{P}_{22}\\ P_{32}\end{array}\end{array}\begin{array}{c}{P}_{03}\\ P_{13}\\ \begin{array}{c}{P}_{23}\\ P_{33}\end{array}\end{array}\right)\ast \left(\begin{array}{c}{r}_{00}\\ r_{10}\\ \begin{array}{c}{r}_{20}\\ 0\end{array}\end{array}\begin{array}{c}{r}_{01}\\ r_{11}\\ \begin{array}{c}{r}_{21}\\ 0\end{array}\end{array}\begin{array}{c}{r}_{02}\\ r_{12}\\ \begin{array}{c}{r}_{22}\\ 0\end{array}\end{array}\begin{array}{c}{r}_{03}\\ r_{13}\\ \begin{array}{c}{r}_{23}\\ 1\end{array}\end{array}\right)\ast \left(\begin{array}{c}{t}_{00}\\ t_{10}\\ \begin{array}{c}{t}_{20}\\ 0\end{array}\end{array}\begin{array}{c}{t}_{01}\\ t_{11}\\ \begin{array}{c}{t}_{21}\\ 0\end{array}\end{array}\begin{array}{c}{t}_{02}\\ t_{12}\\ \begin{array}{c}{t}_{22}\\ 0\end{array}\end{array}\begin{array}{c}{t}_{03}\\ t_{13}\\ \begin{array}{c}{t}_{23}\\ 1\end{array}\end{array}\right)\ast \left(\begin{array}{c}x\\ y\\ \begin{array}{c}z\\ 1\end{array}\end{array}\right)$$

According to Eq. (16), points of PCD can be mapped to image coordinates through the conversion matrix between lidar and camera. The image with YOLO network recognition is taken as the input. The point cloud can be mapped to the image with the target bounding box. Pixel points in the bounding box are justified through coordinate transformation and mapped to the points in the corresponding PCD. Afterwards, the point cloud coordinates mapped to the image in the bounding box are judged. The point cloud coordinates corresponding to the pixels in the bounding box are counted, which is used to calculate the three-dimensional bounding box of a pedestrian in the PCD. Thus, the position of a target pedestrian in the point cloud is detected with a higher recognition rate.

K-means clustering of point cloud pedestrian

The K-means algorithm is a relatively simple unsupervised clustering approach (Jinghui, Wei & Ru, 2020). The algorithm’s input is a sample set divided into K subsets. K initial cluster centres are randomly selected, and then the distance between each sample and the cluster centre is calculated. The cluster results are output when the partition of the subsets satisfies certain requirements. In this article, the PCD is used as the input, Euclidean distance is utilized as the clustering condition, and the pedestrian recognition results of two-dimensional images are employed as a priori knowledge to cluster the pedestrians satisfying the requirements of clustering in the point cloud scene. The K-means clustering is implemented in the point cloud.

For a given sample set divided into K clusters according to the size of the sample point distance, the set of clusters is denoted by $(C_1,C_2,\cdots \cdots C_k)$, and the iterative method is used to minimize the squared error defined by Eq. (17)

(17) $$E={\sum }_{i=1}^k{\sum }_{X=C_i}{\Vert x-u_i\Vert }_2^2$$

(18) $$u_i=\frac{1}{\left|C_i\right|}{\sum }_{x\in C_i}x$$where $u_i$ denotes the mean vector of $C_i$ and is denoted by Eq. (18). The distance from the point to all centre points belonging to the clusters of the nearest center point is calculated for each point. After running iterative calculations, the centre point of the cluster is recalculated, and the centre point nearest to each point is found until there is no change in the previous two iteration results.

In K-means clustering, determining the initial mean vector for K clusters and selecting the location will affect the pedestrian clustering results and running time (Wang, Gao & Feng, 2020). The conventional K-means clustering selects the mean vector completely at random, which will slow the convergence speed of the algorithm. Therefore, choosing the appropriate mean vector and the number of clusters, K, is necessary. The commonly used methods to determine K include the elbow and contour coefficient methods (Pei & Rui, 2015). The elbow method uses the sum of the square errors (SSE) to select the appropriate K defined by Eq. (19). As K increases, the sample division will become more and more refined, and the aggregation degree of each cluster will gradually increase. The core index of the contour coefficient method is called the Silhouette coefficient (SC), defined by Eq. (20). The contour coefficient determines the nearest cluster.

(19) $$SSE={\sum }_{i=1}^k{\sum }_{p\in C_i}{\left|p-m_i\right|}^2$$

(20) $$S=\frac{b-a}{max(a,b)}$$

When K is determined by elbow and contour coefficient methods, there will be a clustering error of the pedestrian target in the point cloud. To overcome the problem of pedestrian misrecognition caused by the point cloud clustering, the K-means clustering is optimized by using prior knowledge of image recognition. First, the input PCD is obtained through image mapping. The two-dimensional recognition of prior coordinates for the data set is carried out. The number of pedestrians in the image recognition is taken as K through cross-validation. Then, to reduce the computational complexity, the mean vectors of K clusters are determined by mapping the midpoint of the two-dimensional boundary box (Xin & Jingying, 2020). Because it is uncertain whether the bounding box’s centre point pixel corresponds to the PCD, the PCD corresponding to the pixel in the bounding box is randomly selected as the mean clustering vector.

K and the mean vectors of K clusters are determined. The pedestrian clustering algorithm based on K-means clustering is presented as follows:

(1) After mapping the image to the point cloud, determine the input PCD $\mathrm{D}=\{{\mathrm{x}}_1,{\mathrm{x}}_2,\dots {\mathrm{x}}_{\mathrm{m}}\}$ $\mathrm{D}=\{{\mathrm{x}}_1,{\mathrm{x}}_2,\dots {\mathrm{x}}_{\mathrm{m}}\}$, and use the K-means clustering to classify pedestrians on the point cloud.

(2) Select the appropriate K. The prior knowledge of image recognition is used to select the number of final bounding boxes in the pedestrian recognition of the image as K.

(3) The centre of clustering is set by selecting K mean vectors $\left\{{\mathrm{u}}_1,{\mathrm{u}}_2,\dots {\mathrm{u}}_{\mathrm{k}}\right\}$ $\left\{{\mathrm{u}}_1,{\mathrm{u}}_2,\dots {\mathrm{u}}_{\mathrm{k}}\right\}$ from the PCD. The points in the bounding image box are mapped as point clouds as the mean vectors of clusters are selected randomly, and one mapping pixel point in each bounding box is selected.

(4) The PCD is iteratively divided and initialized to $\mathrm{C}=\{{\mathrm{C}}_1,{\mathrm{C}}_2,\dots {\mathrm{C}}_{\mathrm{k}}\}$ $\mathrm{C}=\{{\mathrm{C}}_1,{\mathrm{C}}_2,\dots {\mathrm{C}}_{\mathrm{k}}\}$, and the distance between the sample point ${\mathrm{x}}_{\mathrm{i}}(\mathrm{i}=1,2,3,\dots \mathrm{m}\}$ ${\mathrm{x}}_{\mathrm{i}}(\mathrm{i}=1,2,3,\dots \mathrm{m}\}$ and the initial mean vector ${\mathrm{u}}_{\mathrm{j}}\{\mathrm{j}=1,2,3,\dots ,\mathrm{k}\}$ ${\mathrm{u}}_{\mathrm{j}}\{\mathrm{j}=1,2,3,\dots ,\mathrm{k}\}$ of the input point cloud is calculated by Eq. (21).

(21) $$D_{ij}={\Vert x_i-u_j\Vert }_2^2{D}_{ij}={\Vert x_i-u_j\Vert }_2^2$$

Put ${\mathrm{x}}_{\mathrm{i}}$ ${\mathrm{x}}_{\mathrm{i}}$ Then, update the cluster into the category where the distance becomes the smallest with the mean vector.

(22) $${\text{C}}_{{\lambda }_{\text{i}}=}{\text{C}}_{{\lambda }_{\text{i}}}{}^{\cup }\left\{{\text{x}}_{\text{i}}\right\}{\text{C}}_{{\lambda }_{\text{i}}=}{\text{C}}_{{\lambda }_{\text{i}}}{}^{\cup }\left\{{\text{x}}_{\text{i}}\right\}$$

(5) Recalculate the mean vector for all points in ${\mathrm{C}}_{\mathrm{j}}$, as shown in Eq. (23).

(23) $${\text{u}}_{\text{j}}=\frac{1}{|{\text{C}}_{\text{j}}|}{\displaystyle {\sum }_{\text{x}\in {\text{C}}_{\text{j}}}\text{x}}{\text{u}}_{\text{j}}=\frac{1}{|{\text{C}}_{\text{j}}|}{\displaystyle {\sum }_{\text{x}\in {\text{C}}_{\text{j}}}\text{x}}$$

(6) Judge whether all mean vectors have changed. If there is no change, output the clustered point cloud $\mathrm{C}=\{{\mathrm{C}}_1,{\mathrm{C}}_2,\dots {\mathrm{C}}_{\mathrm{k}}\}$.

Basic idea of the compression method

The PCD of three groups of pedestrian scenes on the road in the KITTI data set (format: pcd file, file name: 003922. pcd, 003955. pcd, 003971. pcd) were used as the experimental objects. The three data groups were named pedestrian detection scene on road 1, pedestrian detection scene on road 2, and pedestrian detection scene on road 3.

The PCD used in the experiment contains rich feature information, including both flat points with slow curvature change and feature points with special visual effects. In the experimentation, the normal vector and curvature of the point cloud are calculated, and then the boundary lines and sharp points are extracted. The process of grid reduction on the PCD of other regions is carried out. Finally, the boundary, sharp, and processed points of the grid reduction are fused. Moreover, the duplicate data is deleted. The entire experiment is completed by saving the compression results.

The algorithm consists of two modules: the compression and recognition modules. The basic idea of the compression module is as follows:

Algorithm (Compression)

1. Read the point cloud, and establish the topological structure for the point cloud.

2. Calculate the normal vector and curvature of the point cloud.

3. Extract the feature points and sharp points of the point cloud.

4. Use the nearest neighbour point of the centre point in the grid to replace other points in the flat area and simplify the grid processing. The PCD in the flat area is obtained when calculating sharp points.

5. Fuse the PCD and delete the same points.

6. Save the compressed PCD.

The basic idea of the recognition module includes three parts:

Algorithm (Recognition)

1. Use YOLO to recognize pedestrians’ images.

2. Construct a mapping between PCD and image,

3. Employ K-means clustering to detect targets.

YOLOv5 is used as the basic network for image recognition. The pixel points in the bounding box of the target image corresponding to the PCD to narrow the range of the point cloud target. The output of the pedestrian target bounding box filters the pedestrian in the point cloud (Zanfeng, 2012). After conducting image-to-point cloud mapping, there would inevitably be interference data in the PCD, so it is necessary to process the mapped PCD. The K-means clusters the mapped point cloud to identify pedestrians in the PCD.

The experimentation for the attribute compression of the point cloud is carried out. Using the platform tmc13, the data set’s three groups of test sequences are tested separately. After configuring the environment, the MPEG-based codec software is compiled and debugged. The environment configuration of the experimental platform is shown in Table 1.

The processing of the experimental compression

The results of the experimental compression are as follows:

Table 2 shows the change in the number of point clouds in each process and the final compression rate and the number of boundary points in the point cloud model, which is the smallest. Because of the many surface features of the point cloud model tested in this article, the number of sharp points extracted also grows larger. Extracting these sharp points can ensure that the geometric features of the point cloud are not easily lost while achieving a higher compression rate. To verify the reliability of the compression algorithm in this article, the random sampling method is used for comparative experimental analysis by using the same compression rate. The compression rates used in the experiments for the three groups of PCD are 37.95%, 44.32%, and 36.40%, respectively.

Comparative analysis of the evaluation of the experimental surface area

The total area of the triangulation constructed by the point cloud before and after compression is calculated. The change in the entire area is computed, and the change of the point cloud model features before and after compression is judged. The larger the change rate, the more feature points are deleted during compression. On the contrary, most of the deleted data are redundant, and the model’s features have little change.

The area of the PCD of the three groups based on the experimental objects is computed. The area of pedestrian detection scene I before compression is 27,635.62 cm², the area of pedestrian detection scene 2 before compression is 51,008.45 cm², and the area of pedestrian detection scene III before compression is 62,835.28 cm². Table 3 lists the changes in the area.

Table 4 shows that the compression rate is the same, and the point cloud surface area change rate after compression is the smallest among the three compression methods. The value itself is relatively small, achieving the goal of a higher compression rate and smaller accuracy loss. During the compression process, not many feature points are not deleted, and the geometric features of the point cloud model are well preserved.

Evaluation index of point cloud compression experiment

For the compression of point cloud attributes, the compression mode is called losss-geom-lossy-attrs. This article will quantitatively analyze the experimental results by using the mean square error (MSE) and peak signal-to-noise ratio (PSNR) indicators. The content of the point cloud compression method based on boundary extraction will be directly compared with the experimental results of the corresponding literature. However, the method evaluating the quality of the point cloud will not be used. PSNR is usually used in voice and image processing. PCD, as a new media form, is also applicable. MSE and PSNR are used to evaluate the processing results of the point cloud.

(24) $$PSN{R}_{A,B}=10{\log }_{10}\left(\frac{p_s}{d_{A,B}^{MSE}}\right)$$

(25) $$PSNR=max(PSN{R}_{A,B},PSN{R}_{B,A})$$

$p_s$ denotes the signal peak. The signal peak is expressed by the length of the diagonal, and the point coordinates are normalized within the range [0,1] in Eq. (24).

(26) $$LD={\Vert \left(x_{max},y_{max},z_{max}\right)-\left(x_{min},y_{min},z_{min}\right)\Vert }_2$$

The point cloud is voxelated, and the peak value of each coordinate is expressed in depth precision b bits:

(27) $$p_c=2^b-1$$

The signal peak in Eq. (24) is defined by

(28) $$p_s=\sqrt{3}{p}_c$$

Equation (29) represents the distance from one point to another point. Specifically, it is the square of the coordinate distance of point i in point cloud A (original point cloud) and the coordinate distance of point j in point cloud B (distorted point cloud) that is closest to point i (usually understood as the corresponding point- the modulus of vector (i, j)).

Equation (30) represents the distance from the point to the plane, that is, the length of the projection of the error vector (i, j) in the direction for the lower surface of a normal vector of the j point in the B point cloud, which is also the same value, making the PSNR of the point cloud have a different direction (A to B or B to A).

Finally, the PSNR value of the point cloud can be obtained by substituting the calculation results into Eqs. (24) and (25).

The PCD before and after filtering is represented by $\mathrm{P}$ $\mathrm{P}$ and ${\mathrm{P}}^{\ast }$ ${\mathrm{P}}^{\ast }$ respectively, the three-dimensional coordinate values of $\mathrm{P}$ $\mathrm{P}$ and ${\mathrm{P}}^{\ast }$ ${\mathrm{P}}^{\ast }$ are represented by $\left({\mathrm{x}}_{\mathrm{P}},{\mathrm{y}}_{\mathrm{P}},{\mathrm{z}}_{\mathrm{P}}\right)$ $\left({\mathrm{x}}_{\mathrm{P}},{\mathrm{y}}_{\mathrm{P}},{\mathrm{z}}_{\mathrm{P}}\right)$, $\left({\mathrm{x}}_{{\mathrm{P}}^{\ast }},{\mathrm{y}}_{{\mathrm{P}}^{\ast }},{\mathrm{z}}_{{\mathrm{P}}^{\ast }}\right)$ $\left({\mathrm{x}}_{{\mathrm{P}}^{\ast }},{\mathrm{y}}_{{\mathrm{P}}^{\ast }},{\mathrm{z}}_{{\mathrm{P}}^{\ast }}\right)$, and the total number of surface vertices is denoted by $|\left\{\mathrm{P}\right\}|$ $|\left\{\mathrm{P}\right\}|$.

(29) $$d_{A,B}^{P_{0}2P_0}={\Vert \overrightarrow{e}(i,j)\Vert }_2^2$$

(30) $$d_{A,B}^{P_{0}2Pl}={\Vert \hat{e}(i,j)\Vert }_2^2=(\overrightarrow{e}\left(i,j\right).\overrightarrow{n_j}{)}^2$$

The comparison of mean square error (MSE) and peak signal-to-noise ratio (PSNR) of two methods for the 1st pedestrian detection scenarios on the road is shown in Figs. 1A and 1B.

The comparison of mean square error (MSE) and peak signal-to-noise ratio (PSNR) of two methods for the 2nd pedestrian detection scenarios on the road is shown in Figs. 2A and 2B.

The comparison of mean square error (MSE) and peak signal-to-noise ratio (PSNR) of two methods for the 3rd pedestrian detection scenario on the road is shown in Fig. 3.

The smaller the MSE, the better the point cloud compression effect, and the larger the PSNR, indicating that the smaller the distortion of the point cloud model, the higher the matching degree with the original point cloud model, and the higher the overall compression quality of the point cloud. The results indicate that when the compression rate exceeds 30%, the proposed method performs better than the other method. When the compression rate is 90%, the advantage reaches the maximum, about 5.02% higher than the random sampling algorithm.

The MSE of the proposed method is smaller than that of the other method, regardless of the compression ratio. The proposed method is also superior to the other method when PSNR is used. When the compression ratio is 90%, the advantage grows larger, indicating that the proposed algorithm has certain advantages over other algorithms.

Comparison of the accuracy of image recognition algorithm after point cloud compression

To verify the effectiveness of the proposed algorithm in point cloud pedestrian recognition, the point cloud pedestrian recognition algorithm using voxel division, such as point pillars, and the image-based pedestrian recognition algorithm YOLO are compared. The outcomes are shown in Figs. 4–6, respectively.

First, the algorithms that take PCD and images as input are compared. The point-pillars method uses cylindrical elements to divide PCD and uses 2D convolution instead of a 3D convolution network. The speed of point cloud target recognition is enhanced, and the problem of long-distance pedestrian target recognition is resolved to a certain extent. Based on the mapping image to point cloud target recognition algorithm, the pedestrian target is recognized in the image data, which can recognize the pedestrian target at a relatively long distance. Then, the target object point cloud is quickly filtered through the mapping matrix between the PCD and the image. Through the optimization of the K-means clustering and the proposed algorithm, the target in the point cloud scene could be quickly determined by selecting the target and calculating the bounding box through the image recognition algorithm. The proposed algorithm inputs images and PCD and compresses them before doing the recognition stage, which can realize the recognition of pedestrians in the point cloud scene. However, since both PCD and compression processing are introduced, its accuracy may deteriorate in mapping point cloud images, and the recognition accuracy would become slightly lower than that of the image recognition algorithm.

However, introducing point cloud scenes breaks through the two-dimensional limit and has a wider application range than the simple image recognition algorithm. The experimental results show that the algorithm proposed in the manuscript is only inferior to the image-based pedestrian recognition method and is significantly superior to the above pedestrian recognition method based on PCD. In addition, this article uses the calibration matrix of PCD and image to filter the target point cloud, which largely inherits the accuracy of the image recognition algorithm and has advantages in pedestrian recognition of point cloud scenes.