Study On Key Technology Of Environmental Perception Oriented To Road Traffic

Autonomous driving technology can prevent traffic accidents,reduce traffic jams,improve traffic efficiency,and the control of the car is absolutely hand off.Therefore,the autonomous driving technology has attracted wide attention.An autonomous driving vehicle is a complex system equipped with features for perception,decision-making,and control.The perception module is an extremely important component of autonomous driving technology that employs cameras,light detection and ranging(Li DAR)system,millimeter-wave radar,and other sensors on the vehicle to sense the surrounding road environment and determine precise information about their position and location,size,and direction of movement of surrounding objects,thus ensuring the vehicle is safely and smoothly driving on the road.In terms of traffic road perception,most of the traditional methods only focus on simple road scenes but find it difficult to handle the complex environments,such as the fuzzy edges of roads,object occlusion,and light variation,etc.Moreover,with autonomous driving technology,it is difficult to detect the inherent abstract semantic features of objects based on the shallow artificial design features,which results in the limited performance of detection and the inability to accurately recognize road signs and surrounding objects.Compared with the traditional artificial features,the deep learning features obtained through multi-layer neural networks in the background of big data are better able to handle complex environments,as they are highly abstract.This dissertation aims to integrate deep learning technology with the perception module of autonomous driving vehicles with a focus on lane marking detection and three-dimensional(3D)object detection to achieve the purpose of accurately perceiving its own position and surrounding 3D information.The main contents of this dissertation are as follows:(1)To obtain high-precision lane markings in autonomous driving scenes,we propose a novel lane marking detection method based on structural analysis.Inverse perspective mapping(IPM)and image filtering are utilized in the method to reduce the accuracy loss caused by perspective effects in the scene image and enhance the lane marking area.And a designed convolutional neural network is employed to accurately detect the local lane marking area,and recognizes lane markings according to the geometric continuity of lane marking in space,thereby improving the detection accuracy of lane markings.(2)For the limitation of IPM and the inefficiency of post-processing based on density clustering,we propose a deviation regression guided lane marking detection method without IPM.Different weights to lane markings at different distances are assigned in the method to better improve the prediction of their local offset,and constructs each lane marking instance on account of density clustering by outputting the offset relationship between adjacent positions of lane marking through the network.This method can result in the accurate detection of the lane markings.(3)To solve the problem of reduced detection of current single-stage 3D object detection algorithms that lack shape priors compared to that of two-stage detectors,we propose a 3D object detection method based on the attention mechanism.The 3D voxel features are applied in the method to characterize the Li DAR point cloud,a pyramid network is utilized to detect features of objects of different scales,and then a predefined area based on the top view attention mechanism is simulated to encode the shape of objects to improve the performance of single-stage 3D object detectors.(4)The previous work is limited by the infrastructure of the 3D detection network;therefore,it is impossible to introduce the 3D shape priors to the network and achieve improved detection.To solve this problem,we propose a 3D object detection method based on 3D shape attention.Uniform sampling and an overlapping mechanism are employed in the method to compensate for the sparseness and non-uniformity of the Li DAR point cloud data.The 3D sparse convolution layer is modified to alleviate the problem of loss in height detection experienced in the previous work.By using the two designed attention modules of top view and height direction to obtain 3D shape feature coding,the performance of 3D object detection will improve significantly.