Research And Implementation Of UAV Binocular Vision Depth Perception Technology

UAVs are used more and more widely in aerial photography,agricultural and forestry plant protection,power line inspection,and security.It is of great significance to apply binocular stereo vision to drones to sense the depth information of environment in real time.Stereo matching in binocular vision has been a hot topic of research.The difficulty lies in occlusion,low texture,and repeated textures.The stereo matching algorithm can be divided into local algorithm and global algorithm.The global algorithm has better matching accuracy than the local algorithm,but the time complexity is too high to achieve real-time performance.Although the local algorithm consumes less time,the matching accuracy is not good.This thesis mainly studies the three parts of cost matching,false matching detection and disparity optimization in stereo matching,and optimizes the stereo matching algorithm based on the characteristics of the embedded platform.The main research content of this thesis is as follows:1.For the traditional cost aggregation with fixed-window algorithm,there is a large number of disparity noises in the small window size and the problem of blurring the disparity edges under the large window size.The non-local cost aggregation algorithm is studied.The algorithm extends the support window to the entire map.The edge of disparity is well preserved,but in the process of the algorithm,the dependencies between the data are serious and it is difficult to achieve real-time performance.The method of multi-scale cost-volume fusion is applied to the fixed window cost aggregation algorithm,which reduces the discrete disparity noise and the false matching of low texture areas in the small window cost aggregation algorithm.The experimental results show that the fixed window aggregation algorithm under multi-scale cost-volume fusion performs better in terms of real-time performance and matching accuracy,and has the characteristics of parallel acceleration and optimization.2.The left and right consistency detection algorithms,the uniqueness detection algorithm and the connected region detection algorithm are studied.These algorithms can detect the false matching of the occlusion region effectively,the false matching in the low texture area effectively,and the discrete unsmooth disparity noise effectively.The experimental results show that the fusion of the three algorithms into multistage false matching detection algorithm greatly improves the detection rate of false matching.3.The disparity optimization algorithm based on the weighted median filter and the superpixel segmentation optimization algorithm based on the regional voting are studied.The problem that the two algorithms have large false matching in the occlusion area correction is studied.The nearest neighbor effective disparity filling algorithm is discussed.These two algorithms are combined to effectively improve the false matching in occlusion areas and improve the matching rate.4.Based on the embedded platform NVIDIA Jetson TX2,the overall framework of CUDA parallel processing is studied.The stereo matching algorithm is transplanted to the platform,and the cost calculation,cost aggregation,cost volume fusion and disparity selection are optimized for parallel optimization.Experimental results show that the algorithm can achieve near real time effect.