Research On Real-Time Stereo Vision And Obstacle Detection

Stereo vision system has been widely used in unmanned vehicles and autonomous robots as the advantages of simple structure,low cost,lower power consumption and more dense reconstruction results.With the stereo vision as the core,this paper gives a real-time stereo vision solution for road traffic environment navigation.The main work is as follows:1.A new feature selection based Census matching operator is proposed.This method combines the core idea of feature selection.The evaluation functions such as information gain,correlation and matching error rate are used jointly,then the census matching operator is obtained by forward search and wrapper selection.Experiments show that compared to the normal Census,the improved operator has the advantages of lower error rate and higher efficiency,which is helpful for the algorithm acceleration and hardware deployment.2.Design and implement a FPGA based stereo vision system.Using the proposed feature selection based Census and SGM of weighted path accumulation based on orientation gradient,this paper designs and optimizes a stereo matching algorithm for FPGA,then the design of stereo vision system based on FPGA is introduced in detail.Experiments show that the proposed algorithm has better effect in FPGA based algorithms with good speed and resource consumption.3.A free space estimation and obstacles detection algorithm based on disparity map is proposed.Starting from the principle of binocular geometry,this paper proposes an efficient nonparametric method for road disparity estimation using formula deduction and image analysis.The new algorithm can adapt to the distant undulating road,and reduce miss rate and false detection.We also exploit the texture gradient information to compute free space and detect obstacles.Only the local gradient extreme point is selected as the candidate node to participate in the dynamic programming process,which greatly reduces the amount of calculation.Moreover,the gradient information is included in the cost equation,which helps to obtained more accurate results.Experiments in several scenes show that,the proposed algorithm outperforms the standard methods both in terms of detection accuracy and runtime performance.With high run-time performance,the algorithm is beneficial to following tasks such as autonomous navigation and object recognition.