Research On Autonomous Driving Technology Based On Multi-mode Composite Navigatio

Visual perception is currently a hot spot for autonomous driving research,which is mainly focused on structured roads with relatively little research on unstructured roads.The working environment of unmanned vehicles contains both structured and unstructured roads,and the road conditions are relatively harsh.Therefore,in this thesis,the autonomous driving technology for unmanned vehicles on structured and unstructured roads is investigated and validated on the RIA-R100 mobile platform.The main work of this thesis is as follows.(1)Aiming at the poor robustness of traditional lane line detection methods,campus road dataset and CULane dataset were used to train Lane ATT network,and the F1 score of test set reached 0.8729,which can effectively detect lane lines under complex situations such as defacement,occlusion,and nighttime.The global path planning was completed with Baidu Map Web API,and along the global path the unmanned vehicle was controlled by proportional differential controller based on the lane line detection results in combination with GPS information.(2)Aiming at the roads without high precision maps,the improved direction evaluation function Dynamic Window Approach(DWA)was proposed.And by introducing the unmanned vehicle orientation and global path point,the dynamic obstacle avoidance and steering function of the local path planning of the unmanned vehicle was implemented,so that the unmanned vehicle can better recover the vision control after completing the local path planning.(3)Aiming at the problem that the mean intersection over union(m Io U)of the training results of unstructured road semantic segmentation dataset is relatively low,a higher level of semantics was used to reclassify the RUGD dataset according to passability.HRNet+OCR network was trained,and the m Io U of the test set reached 80.88%.Since the unstructured road boundary is ambiguous,for the unstructured road boundary extraction and for the problem of large computation of Random Sample Consensus(RANSAC)algorithm,RANSAC algorithm was improved and the parameters were adaptively updated,which effectively reduces the number of iterations and reduces the computation time by about 50%.The improved RANSAC algorithm speeds up unstructured road boundary extraction.(4)Aiming at the problem that the lidar obscured on unstructured roads,a local path planning method based on the segmentation map evaluation function was proposed to implement the obstacle avoidance function.For the problem of bumps on the unstructured roads,a dynamic inverse perspective transformation method based on the change of pitch angle was adopted to real-time recover the parallelism of the road.A monocular ranging method based on YOLO detector was used to improve the safety of unmanned vehicles in autonomous driving on unstructured road.(5)The corresponding software was designed and implemented based on ROS.Autonomous driving functions of unmanned vehicles on structured roads and unstructured roads were implemented.The simulation test and practical test of structured road and unstructured road on campus were completed,which proved the validity of the proposed methods.