Robust Stereo Visual Odometry Based On Sparse Features

SLAM(Simultaneous Localization and Mapping)is a classical problem in the field of robots.The SLAM system based on stereo vision has the higher degree of concern in the SLAM study because the scene is widely used and the advantages of no scale drift.Visual odometry is the front-end module of SLAM system and is responsible for receiving sensor data and calculating poses of adjacent frames.It has an important influence on the performance of SLAM system.This paper constructs a stereo visual odometry based on sparse features.The algorithm is aiming at improving the existing problems of visual odometry,in order to promote the visual odometry accuracy and real-time improvement.This paper is mainly from the extraction of feature points,the left and right image matching,the matching between frames,the pose solution is improved.The main work of this paper includes the following five aspects.First,proposing uniform feature points' algorithm based on image Pyramid segment.In order to overcome the shortcomings that the features get together due to the global response value.Using image segmentation and Pyramid based on the local response value,the homogenization of feature points.Second,proposing the distance point matching algorithm based on adaptive threshold.According to the scope of the search about the fixed threshold by using graph matching and matching leads to uncertain problems.Using the firstly after matching feature point information,calculating the search threshold of each feature point and matching two times,two.Thirdly,proposing matching algorithm based on the adaptive window under the assumption of uniform motion.The matching outliers are too much for too little a priori information between frames matching.Gradually increasing the prior information,and designing based on the distance calculation of feature point search function,improving the accuracy of the matching.By experiment shows that comparing the algorithm and the original matching,there is great improvement effect.Fourth,proposing pose solution algorithm based on multiple constrains and deducing equation.According to the matching result,pushed to the residual function,Jacobian matrix,the final state will increment equations of each constraint are simultaneously.The final solution more accurate than a single constraint pose.Fifth,Improving the Huber kernel function based on adaptive threshold.The problem of inaccurate point for Huber kernel function using a single threshold value of all points which confirmed that the proposed adaptive threshold function based on the distance,do for each feature point can be Huber proper threshold,makes the Huber function in the residual processing,more accurately distinguish internal and external point.In summary,this paper improves the performance of feature extraction,feature matching and pose solution.It improves the accuracy of visual odometry.The proposed algorithm through the KITTI dataset odometry test,and get better results.