Development Of Lane Keeping And Obstacle Avoidance System Based On Multi-sensor Information Fusion

Driven by the artificial intelligence wave,the automotive industry is gradually developing towards the direction of intelligence and modernization.Advanced auxiliary driving system has become an integral part of modern cars and plays an important driving role in the upgrading and transformation of new vehicles.Lane holding and obstacle avoidance system is the core of auxiliary driving system,and integration and upgrading with emerging technology is the main direction of current research.This subject mainly focused on the application of multi-sensor information fusion algorithm in lane holding and obstacle avoidance system.The two sensor detection features of camera and lidar were fused to overcome the shortcoming that a single sensor could only sense one feature of the target.During the driving process,the vehicle could make more accurate judgment based on multi-dimensional information and realize the vehicle driving independently in the set scene.The whole system involved many technical fields.Combining with relevant research at home and abroad and existing technology,the subject was mainly realized through the following parts:(1)In this paper,Hough transform algorithm and LBP feature cascade classifier were used for lane line detection.The image frame was preprocessed in Open CV image processing framework.After removing the interference information in the image,edge points were extracted by using Canny edge detection operator,and the feature points were remapping to Hough space to get the line information.LBP feature was used to train SVM classifier to supplement Hough detection algorithm and make up for the shortcoming of Hough algorithm in detecting discontinuous lane lines.(2)In this paper,the Mobile Net-SSD object detection network was used to detect obstacles against,which could realize real-time detection of obstacle position after deployment on the mobile terminal and acceleration with Open VINO.In the study of the depth calculation of obstacles,radius filtering was used to remove the outlier noise points in 2D point cloud,DBSCAN clustering algorithm was used to classify the point cloud data,and depth settlement method based on corner distance weighting was used to measure the depth of obstacles.(3)The output results of the target detection network and the target depth information were jointly labeled,the neural network framework was designed,and the cumulative error gradient descent method was used to train the network to realize the mapping of the two characteristic information to high latitudes.A graphical interface test system was designed using MFC visual programming.Each part of the system was tested separately.Finally,it was tested and analyzed with hardware part.