Research On Stereo Vision-based Obstacle Detection Of Unmanned Vehicle

Unmanned vehicle technology is one of the research hotspots in today’s cutting-edge science and technology,it is for many aspects of social and economic development are of great influence.Unmanned vehicle technology involves several problems,include vehicle navigation,vehicle communication,active obstacle avoidance,and autonomous cruise.Its central problem is obstacle detection.One of primary technology of obstacle detection is stereo vision-based obstacle detection technology,containing stereo image acquisition,stereo matching and obstacle recognition of three aspects.Focusing on stereo matching and obstacle recognition,the author had done amount of researches on these two aspects,analyzing and discussing the corresponding technical method.The followings are the main researching content in this paper.(1)Focusing on unmanned vehicle scenes,the stereo matching technology based on convolution neural network has been studied.On the basis of analysis of the convolutional neural network structure and corresponding stereo matching technology,the Asymmetric Kernel-Siamese Convolutional Neural Network(AK-SCNN)is designed,which improve the structure of network.The matching cost is calculated with AK-SCNN,and finally the disparity is computed.Through experimental analysis,the best hyperparameters of AK-SCNN such as the setting of network structure is selected.Comparing with other traditional local stereo matching algorithms,it has lower bad matching rate.(2)The obstacle detection method using V-disparity map is studied.Firstly,the principle of V-disparity and the feature of surface projection in V-disparity map are studied.Then,on the basis of analysis of the traditional V-disparity method,and according to maximum and minimum constraint and distance constraint,the road line adaptive threshold extraction algorithm is proposed,which can improve the obstacle detection method using traditional V-disparity map.The experimental results show that the improved method can effectively detect the obstacle area.Comparing with the traditional method,it can raise the recall and accuracy rate of obstacle detection.(3)The obstacle detection method using point cloud map is studied.Firstly,the acquisition and simplification of point cloud map,the definition of point cloud density and the rasterization method are studied.Then,focusing on the road environment,the road ideal point cloud density fitting method is proposed to fit the ideal point clouddensity of road.According to compare real point cloud density with ideal point cloud density,the obstacle area is detected.The experimental results show that this method can succeed to divide obstacles area,and prove the feasibility of this method.Comparing with V-disparity method,the experimental results show that the two method have their own advantages and disadvantages,which are suitable for different scenarios.