Research On Track Recognition Method Based Onfusion Of Lidar And Visual Information

With the rapid development of information physical system and information communication technology,it is an inevitable trend for vehicles to realize automation,intelligence and Internet.The formula auto driving car provides a platform for the development and validation of new technologies under challenging conditions.It provides a unique opportunity to test the hardware,software,algorithm and risk index required by the auto driving vehicle,and accumulates technology and experience for the commercial and popularization of auto driving.This paper focuses on the obstacle detection and camera image recognition algorithm of lidar based on the environment perception technology of formula auto driving car.In order to provide necessary and accurate information for the path planning and vehicle control of the formula auto driving car and realize the safe driving of the car,the difficulties of target detection,segmentation,recognition and positioning in the three-dimensional environment of the formula auto driving track for college students are overcome.Firstly,an obstacle detection algorithm based on 3D lidar is designed.The installation position of the lidar is installed in the center of the front wing of the car to ensure a large number of relatively clean radar point cloud data.Then,in order to facilitate the subsequent point cloud processing,the k-d tree is used to construct the topological relationship of the three-dimensional point cloud,and the guided filtering method based on statistical filtering is used to smooth and simplify the point cloud,so as to reduce the amount of point cloud data,improve the overall computing speed and save computing resources,and smooth out the noise due to various reasons,Then,ransacs random sampling consistency algorithm is used to segment the plane,conditional Euclidean clustering is used to detect the appropriate size of obstacles,and the coordinates and size of the obstacles are calculated to realize the detection and location of the target obstacles.Secondly,an obstacle classification and color recognition algorithm based on monocular camera and deep learning is designed.Firstly,the monocular camera is installed above the lidar in the center of the front wing of the car and below the nose wing to ensure good vision and lighting conditions.Using the improved resnet18 network as the benchmark model,using the mish activation function,and introducing the channel attention mechanism,it has the advantages of lightweight and high accuracy.The target image is classified into four categories(non bucket,red bucket,blue bucket and yellow bucket).Finally,a calibration fusion method of lidar and monocular camera is designed.Based on the camera imaging model and coordinate transformation method,the obstacle coordinates and the approximate length,width and height detected by lidar are mapped to the two-dimensional RGB image of the camera,and the target obstacle is segmented from the image,and then the segmented image is sent to the improved residual image classification and color network with channel attention mechanism.In this way,the coordinates,types and colors of obstacles can be obtained.The experimental results show that the proposed track recognition method based on lidar and visual image fusion can locate,classify and color recognize the obstacles in the track environment in 100 ms cycle,the accuracy is within 4 cm,the accuracy of classification and color recognition is 99.63%,and it has good real-time performance,accuracy and robustness.It can provide accurate,real-time and reliable track information for the path planning of formula auto driving car.The results show that the average automatic driving speed of the formula car is 6.3m/s.