Research On Road Scene Modeling And Intelligent Vehicle Motion Planning Based On Monocular Vision

Intelligent vehicles are an important direction for future vehicle development,involving key technologies in multiple fields such as vehicles,computers,and communications,and have become a hot research topic in the automotive field.At present,autonomous driving smart vehicles still have many technical problems in terms of perception,decision-making,planning,etc.,which hinders their industrialization.This paper aims at structured road scenarios,based on monocular vision,from other vehicle position and pose estimation,lane line semantic segmentation,Research on road scene modeling and motion planning,the main research results are as follows:(1)Estimate the position and p osture information of his car based on monocular vision.Select position information,size information and heading angle as the description parameters of his vehicle's posture,according to the constraint relationship between 2D bounding box and 3D bounding box,based on MobileNetv2 neural network to estimate th e 3D bounding box of the vehicle in the image,so as to obtain the vehicle's position,Attitude and size.The 3D bounding box obtained by regression of the target detecting the 2D bounding box can well express the position,size and direction of the target.The positioning error is about 15%,the size error is between 5%and 10%,and the angle estimation error is 12.8%.The effectiveness and accuracy of the algorithm are verified through real vehicle experiments.(2)Lane line detection and road scene modeling.Using ResNet101 as a skeleton network to establish a network model for extracting lane line features in images,the test results on the ApolloScape and KITTI datasets have verified the reliability and robustness,accuracy,precision and Meanlou and other evaluation indicators have reached more than 90%.A method combining DBSCAN clustering algorithm and polynomial is designed to cluster and fit the refined lane lines,while further extracting lane line information,the noise existing in the lane line features and the shadow of the trunk line on the lane line detection zone are eliminated Interference.Finally,combined with the vehicle pose information,the lane line information and the vehicle pose information are unified by establishing pixel weights,and the road scene model is established to more intuitively reflect the true distance and relative relationship between the vehicle and the surrounding environment,based on monocular Vision provides methods and ideas for road scene modeling research.(3)Research on motion planning based on Frenet coordinate system.Based on the reference path,the Frenet coordinate system is established.From the constraints of vehicle kinematics and dynamics such as ride comfort,handling stability and comfort,the horizontal and vertical penalty functions and the total penalty function are designed to complete the horizontal and vertical motion planning of the vehicle.And jointly obtain the final optimal path.Before and after introducing penalties such as lateral,longitudinal speed,and acceleration,the lateral speed,lateral acceleration,and heading angle of the vehicle change smoothly,eliminating sudden changes in longitudinal speed and acceleration,making the vehicle less impacted and more comfortable.The motion planning results in the road scene model established by the road images collected by the experimental vehicles also verify the effectiveness and accuracy of the algorithm.