CN-118587543-B - Three-dimensional point cloud vehicle detection method in expressway scene based on voting voxel fusion network

CN118587543BCN 118587543 BCN118587543 BCN 118587543BCN-118587543-B

Abstract

The invention discloses a three-dimensional point cloud vehicle detection method in an expressway scene based on a voting voxel fusion network, which comprises the following steps of S1, constructing a three-dimensional point cloud vehicle detection dataset KITTI-Car-Weather in the expressway scene, comprising conventional, foggy, rainy and snowy days, S2, constructing an FN-DHV-VDHS network model for three-dimensional point cloud vehicle detection in the expressway scene according to the characteristics of the dataset and the requirements of detection tasks, S3, training the model and optimizing parameters, constructing experimental result evaluation indexes of multiple Weather conditions, and respectively evaluating the detection average precision, the detection speed and the robustness of the three-dimensional point cloud vehicle detection method in the expressway scene based on voxel network fusion depth Hough voting. The performance and the robustness of the model constructed by the invention are superior to those of other commonly used vehicle detection networks, the average accuracy of vehicle detection in a highway scene reaches (93.86,84.49,83.40), and the detection frame rate reaches 20.5Hz. The relative corrosion error under multiple weather conditions is only 73.29%.

Inventors

ZHAO CHIHANG
Xie xingpeng
LI XUAN
DENG WENHAO

Assignees

东南大学

Dates

Publication Date: 20260508
Application Date: 20240606

Claims (7)

1. A three-dimensional point cloud vehicle detection method in a highway scene based on voxel network fusion depth Hough voting is characterized by comprising the following steps: s1, constructing a three-dimensional point cloud vehicle detection dataset KITTI-Car-Weather in a highway scene, wherein the three-dimensional point cloud vehicle detection dataset comprises a conventional day, a foggy day, a rainy day and a snowy day; S2, aiming at the characteristics of a data set and the requirements of detection tasks, constructing an FN-DHV-VDHS network model for detecting the three-dimensional point cloud vehicles in the road scene; s3, model training and parameter optimization are carried out, experimental result evaluation indexes of multiple weather conditions are constructed, and the average detection accuracy, the detection speed and the robustness of the three-dimensional point cloud vehicle detection method in the highway scene based on the voxel network fusion depth Hough voting are respectively evaluated; In the step S2, an FN-DHV-VDHS network model for three-dimensional point cloud vehicle detection in a highway scene is constructed, and a network is constructed by adopting a feature extraction module fusing points and voxels, wherein the specific construction steps are as follows: s21, constructing a feature extraction proposal frame generation network based on voxels: the first part voxelizes the point cloud and extracts original voxel characteristics by utilizing a voxel characteristic coding layer; the second part combines the space sparse convolution and the sub-manifold sparse convolution into a 3D sparse convolution layer, further feature extraction is carried out on voxels, then a bird's eye view is formed by utilizing a spark-to-dense layer, 3D sparse features are converted into dense 2D features, and the third part utilizes a region generating network comprising a 2D backbone network to extract the bird's eye view features and generate candidate regions; s22, constructing a voting point perception network based on the depth hough voting: The method comprises the steps of taking original point clouds as input data, carrying out point sampling by adopting a brand new multi-range sampling method comprising a segmentation-based sampling method, a furthest point sampling method and a characteristic sampling method, firstly carrying out first downsampling by the furthest point sampling to generate two groups of point clouds, secondly carrying out second downsampling by using the characteristic sampling to the two groups of point clouds, carrying out segmentation-based sampling to one group of point clouds and carrying out third downsampling by the other group of point clouds, and finally respectively sampling to a seed point set and a key point set; s23, constructing a vehicle detection network with fusion of points and voxel characteristics: and performing grid division on a proposal frame generated by a voxel-based feature extraction proposal frame generation network by using a RoI-grid pooling module, performing feature aggregation on a point set consisting of voting points and key points by setting variable radiuses with the grid points as centers, and finally performing confidence prediction and proposal frame optimization by using aggregated features.
2. The method according to claim 1, wherein in the step S1, a Weather-corrosion simulation algorithm based on physical principles is used to simulate a scene containing vehicles, using KITTI datasets as an original point cloud, and a three-dimensional point cloud vehicle detection dataset KITTI-Car-Weather in a highway scene is constructed.
3. The method according to claim 2, wherein the specific method for constructing the three-dimensional point cloud vehicle detection dataset KITTI-Car-Weather in the highway scene is as follows: S11, collecting KITTI data sets, removing data with Campus and Person as scenes in the data sets, storing the rest point cloud data containing vehicles as data sets KITTI-Car, and re-labeling by adopting LabelCloud software; S12, dividing the remarked KITTI-Car data set into a three-dimensional point cloud vehicle detection training set and a verification set according to the proportion that 80% is the training set and 20% is the verification set; S13, inputting KITTI-Car data sets into a weather corrosion simulation algorithm to simulate point cloud data corresponding to weather conditions; S14, simulating the Weather corrosion simulation algorithm to obtain point cloud data, dividing the point cloud data into Light, moderate and Heavy according to the corrosion degree, collecting the tidying data and constructing a three-dimensional point cloud vehicle detection data set KITTI-Car-Weather in the expressway scene.
4. A method according to claim 3, wherein for the weather corrosion simulation algorithm, the physical fog day simulation method is used to create fog day data, the LISA rain day simulation method is used to create rain day data, and the LISA snow day simulation method is used to create snow day data.
5. The method of claim 4, wherein the attenuation coefficient of the foggy day Coefficient of backscattering Respectively is And Rainfall rate in rainy days Is that Snowfall rate in snowy days Is that 。
6. The method according to claim 1, wherein the step of constructing experimental result evaluation indexes of multiple weather conditions in the step of S3, respectively evaluating the detection average accuracy, detection speed and robustness of the three-dimensional point cloud vehicle detection method in the road scene based on voxel network fusion depth hough voting, and analyzing the improvement strategy effect of the model, comprises the following specific construction steps: s31, constructing experimental result evaluation indexes under multiple weather conditions: in the evaluation index of vehicle detection, the detection precision is selected to be the average precision AP40@0.7, the detection speed is selected to be the frame rate FPS, and the robustness is selected to be the relative corrosion error RCE; s32, constructing an improved strategy effect comparison analysis: And respectively carrying out comparative analysis on the effectiveness of the fusion detection of the voting algorithm and the effectiveness of the improvement of the voting algorithm under the condition of multiple weather.
7. The method of claim 6, wherein the specific content of the evaluation index of the three-dimensional point cloud vehicle detection experiment result in the road scene with the multi-weather condition constructed in the step S3 is as follows: calculate the detection average accuracy AP40@0.7 (IoU =0.7): Wherein, the The cross-over ratio is represented by the ratio, A bounding box representing the prediction is presented, Representing a real bounding box; the average degree of accuracy is indicated as well, Representing the probability of predicting as actually being a positive sample among the positive samples, Representing the probability of being predicted as a positive sample among the actual positive samples; Calculating a detection speed frame rate FPS: Wherein, the (FRAMES PER seconds) represents the number of frames per second, Representing the time required to detect a frame of point cloud; calculating a robustness relative corrosion error RCE: Wherein, the Indicating the average accuracy of the test under corrosive conditions, Is shown in In such weather conditions The average accuracy of the detection of the severity; Indicating the relative corrosion error, Expressed as the average accuracy of the test under corrosive conditions.

Description

Three-dimensional point cloud vehicle detection method in expressway scene based on voting voxel fusion network Technical Field The invention belongs to the field of intelligent high-speed intelligent perception research, and particularly relates to a three-dimensional point cloud vehicle detection method in a road scene based on a deep learning fusion network. Background The intelligent perception level of the expressway is the key of expressway management, and a reliable and effective 3D vehicle detection algorithm is the basis and guarantee of the intelligent perception level of the expressway. The 3D vehicle detection in the expressway scene is one of the most important tasks for understanding and analyzing the highway environment information, and aims to accurately position the vehicle by utilizing three-dimensional point cloud data in the expressway scene and simultaneously identify the type information of the vehicle. Because the three-dimensional point cloud has the advantages of providing accurate position and geometric information of the target and being insensitive to illumination change, the vehicle detection based on the three-dimensional point cloud target detection can better acquire three-dimensional space characteristics to finish target detection by means of the provided depth information of the point cloud. However, there are many problems in the vehicle detection in the expressway scene, on the one hand, the vehicle has a fast moving speed, and on the other hand, the road environment in spring and autumn has a night foggy day, and the mountain hilly area is particularly obvious. These characteristics enable a 3D vehicle detection model deployed in a real system to meet the requirements of reasoning speed while having high detection accuracy. In addition, the rain and snow weather that often occurs in expressway scenes can lead to a rapid degradation of the performance of the lidar. Such weather changes require that the 3D vehicle detection model be robust. In view of the above, the invention designs a voxel fusion network three-dimensional point cloud vehicle detection model based on depth Hough voting, and solves the problem of balance of three-dimensional point cloud vehicle detection precision, efficiency and robustness under multiple weather conditions in a highway scene. Disclosure of Invention The invention aims to overcome the defects in the prior art, provide a three-dimensional point cloud vehicle detection method in an expressway scene based on deep learning, construct an FN-DHV-VDHS network model for three-dimensional point cloud vehicle detection in the expressway scene, effectively detect the three-dimensional point cloud vehicle in the expressway scene through model design and training optimization, solve the problem of balance of three-dimensional point cloud vehicle detection precision, efficiency and robustness under multiple weather conditions in the expressway scene, and provide a basis for the subsequent deployment of three-dimensional point cloud vehicle detection in an all-weather expressway scene. The technical scheme adopted by the invention is that the three-dimensional point cloud vehicle detection method in the expressway scene based on the voxel network fusion depth Hough voting. The method comprises the following steps: s1, constructing a three-dimensional point cloud vehicle detection dataset KITTI-Car-Weather in a highway scene, wherein the three-dimensional point cloud vehicle detection dataset comprises a conventional day, a foggy day, a rainy day and a snowy day; S2, aiming at the characteristics of a data set and the requirements of detection tasks, constructing an FN-DHV-VDHS network model for detecting the three-dimensional point cloud vehicles in the road scene; S3, model training and parameter optimization are carried out, experimental result evaluation indexes of multiple weather conditions are constructed, and the average detection accuracy, the detection speed and the robustness of the three-dimensional point cloud vehicle detection method in the highway scene based on the voxel network fusion depth Hough voting are respectively evaluated. . Preferably, in the step S1, a KITTI data set is adopted as an original point cloud, a Weather corrosion simulation algorithm based on a physical principle is utilized to simulate a scene containing vehicles, and a three-dimensional point cloud vehicle detection data set KITTI-Car-Weather in a highway scene is constructed, which comprises the following specific steps: S11, collecting KITTI data sets, removing data with Campus and Person as scenes in the data sets, storing the rest point cloud data containing vehicles as data sets KITTI-Car, and re-labeling by adopting LabelCloud software; S12, dividing the remarked KITTI-Car data set into a three-dimensional point cloud vehicle detection training set and a verification set according to the proportion that 80% is the training set and 20% is the verification set