Research On Drivable Area Detection Technology Under Dynamic Circumstances

Intelligent driving technology receives more and more attentions in recent years.It can help vehicles complete the tasks of self driving,path planning,obstacle avoiding and so on in semi-automatic or full-automatic manners.Intelligent driving technology can alleviate traffic congestion,reduce the probability of traffic accidents and improve the social productivity.At present,many companies and scientific research institutions are involved in the field of intelligent driving and have made great achievements.The accuracy and the robustness of intelligent driving have reached a high standard.However,most of the relevant research is carried out on vehicles with integrated hardware platform.Thus,the technology can not be applied on a large scale because of the huge cost of hardware and software.In order to relieve these problems,this paper improves the perception algorithm in intelligent driving based on some relevant researches.A driveable area detection network solely based on visual sensor is proposed.The network consists of the lane detection branch and obstacle detection branch,which can output the lane and the obstacle information,respectively.Two branches are trained separately on the same dataset.In addition to the driveable area detection network,this paper also proposes an area intersection algorithm,which utilizes the proportion of obstacles in the lane area to judge the congestion degree of a lane,then the appropriate driving suggestions are sent to vehicles.The main contents of this paper are concluded as follows:We explore the lane detection network based on the row anchor classification method.The traditional image segmentation algorithm consumes a large amount of calculation because it needs to classify each pixel of the input image.To obtain a better real-time effect,the lane detection network in this paper predicts some preset key points of the image,from which the lane can be inferred.This kind of method is not only accurate,but also can effectively save the calculation cost.In addition,we also make some improvements in the situation where some lanes are invisible,we append the spatial convolution modules and attention modules into the network,they can help better integrate the global information from the input image and thus achieve higher detection accuracy.Finally,we conduct some experiments on the CULane dataset.We explore the obstacle detection network based on one-stage object detection method.Many contemporary object detection algorithms are built for daily life scenes.Thus,they do not fully inspire the specific features of obstacles in driving environment.Therefore,we design a obstacle detection network specially for driving environment.We modify the feature extraction layer of the network to better fit the driving environment.The experiment results show that the modification can help the network improve the accuracy by nearly 4%.Besides,in order to ensure the consistency with the dataset used in lane detection network,we propose a new dataset based on the CULane dataset.The original dataset only contains lane information and we add the obstacle information.We fuse the lane detection network and obstacle detection network to build an integrated driveable area detection network.Two networks share the same feature extraction layer and predict the road information and obstacle formation,respectively.Furthermore,we also propose an area intersection algorithm as the extensive research.First,we calculate the intersection area of lane and obstacles,then we estimate the congestion of road according to the proportion of the intersection area in the lane area,and finally we output the intelligent driving suggestions to the intelligent vehicles.