Research And Realization Of Distance Measurement Method Based On Binocular Vision

With the rapid development of driving assistance systems,robot navigation and virtual reality,spatial location information has attracted more and more attentions.The binocular stereoscopic vision system searches for the position of the target in the left and right images by imitating the human eye vision system,and determines the target distance according to the triangulation principle.Stereo matching is a key factor influencing the accuracy of distance measurement.More and more scholars have put into the research to improve the accuracy of stereo matching.The current stereo matching method includes local algorithm,semi-global algorithm and global algorithm.The local algorithm is simple and fast,but the false matching rate is high in the low texture region.The semi-global algorithm can generate better disparity map better than the local algorithm,but the time is very costly.The global algorithm can generate the most accurate disparity map,but the cost is too high.It is difficult to meet the real-time requirements of binocular distance measurement system.In this thesis,we improved the accuracy of local algorithms and the real-time performance of semi-global algorithms.Moreover,we transplant the improved local algorithms to DSP embedded platform and construct a binocular distance measurement system.The main contents focus on the following aspects:Firstly,aiming at the problem that the localized stereo matching algorithm has high error matching rate in low texture region,a local stereo matching method based on stable point is researched.The disparity map of the stable point is calculated,and the stable point disparity is transmitted to the unstable point.On one hand,the stable grid points are selected according to the similarity of the neighborhood of the grid.On the other hand,the stable edge points are selected by the confidence score.Moreover,the stable grid points and the stable edge points are interpolated to get a dense disparity.Secondly,for the semi-global stereo matching algorithm real-time poor problem,the algorithm improved from three aspects to improve speed.Firstly,the aggregation direction of the algorithm is reduced,and then the aggregation direction combination is optimized.Finally,the minimum number of pixels is calculated by grid sampling.The minimum spanning tree is established based on the original image and the dense disparity map is generated by upsampling the grid disparity map.Thirdly,a disparity correction method based on the superpixel segmentation and the confidence map is researched.The area to be corrected is determined according to the confidence map.Then we correct the disparity map in accordance with the principle of voting.Fourthly,the local stereo matching algorithm based on the stable point is transplanted into the DSP development platform,and the speed of the algorithm is improved after many aspects optimization.Overall,the binocular distance measurement system is constructed and the distance precision is analyzed.On the Middlebury platform,it is proved that the improved algorithm proposed in this thesis is effective,and the accuracy of the local algorithm and the real time of the semi-global algorithm are improved.The real-time binocular distance measurement system on DSP is realized.