Research On Environment Perception Technology Of Autonomous Driving Vehicle Based On Deep Learning

The demand for autonomous driving vehicles is getting higher and higher in today’s society,and the related technology research is also in the process of rapid advancement.As the primary module in the autonomous driving system,environment perception is responsible for providing accurate and reliable information about the surrounding environment to help the vehicle identify the driving area,avoid obstacles and plan the driving path.It is the critical technology in the autonomous driving system.This thesis focuses on the environment perception technology in the autonomous driving system,designs,improves and experimentally verifies the various tasks in the environment perception.Including pedestrian,vehicle,and traffic light detection algorithm.Design and improvement of lane recognition algorithm based on the traditional technique and deep learning method.Design of curb detection algorithm.A method for multi-task fusion of detection and segmentation tasks is proposed,the network model in this method can simultaneously detect obstacles and lane lines.The details are as follows.Firstly,this thesis designs a pedestrian,vehicle,and traffic light detection method.Through the improvement of the network structure and the open dataset,the results show that this method can accurately detect pedestrians,vehicles and traffic lights in real-time performance,and is very robust in detecting small targets and the situation of illumination changesSubsequently,this thesis designs a real-time detection method for lane lines based on traditional image processing technology.With improved vision features and post-processing techniques,a more accurate and real-time lane line detection result can be achieved.Thirdly,this thesis implements a multi-lane-lines recognition algorithm by using deep learning technology.With the scene segmentation model and the improved recognition strategy,an excellent multi-lane-lines recognition result can be achieved.This method can deal with the situation when lane lines are unclear,occluded,or illumination changes.This thesis also analyzed and compared these two kinds of lane line detection methods mentioned above.Then,this thesis also designed a method for curb detection.The LiDAR point cloud is preprocessed by the proposed region of interest method so that the real-time performance of the detection algorithm is improved.Then the data is smoothed by the cleaning algorithm.Finally,multiple curb points are detected by using the sliding window.Combined with the lane line detection algorithm proposed in the thesis,this method can provide real-time and accurate road information to the vehicle through real road testing and verification.Finally,this thesis proposes a multi-tasking network that can perform obstacles detection tasks and lane line recognition tasks simultaneously.With the designed end-to-end fusion network,multi-task loss function,and improved open dataset,this method enables simultaneous detection of multiple obstacles and lane lines.With experimentation,this method can effectively enhance real-time performance and reduce the need for hardware and consumption.