Research On Graph Optimization Algorithm With Loop Closure For Visual SLAM Back-end

Mobile robot technology is one of the most frontier fields in the development of current science and technology.Simultaneous localization and mapping is the first problem encountered by robots in an unknown environment.With the popularity of visual sensors,Visual SLAM has been widely concerned.A visual SLAM system can be partitioned into two parts:front-end visual odometry and back-end optimization.The role of visual odometry is to produce rough pose estimation and map representation;Back-end optimization based on loop closure is very important to eliminate the accumulated error of robot pose estimation and realize the consistency of the map.Loop closure detection and graph optimization are the cornerstones of back-end optimization for SLAM system.Real-time performance is the main factor affecting the application of SLAM system.Loop closure detection technology based on online visual dictionary and COP-SLAM method are the best representatives of real-time performance respectively in loop closure detection and graph optimization.This principle of loop closure detection based on visual dictionary is analyzed.Based on greedy strategy,a real-time loop closure detection approach using online visual dictionary is proposed.Process of dictionary construction gives priority to dealing with SURF feature that has the maximum Euclidean distance from the closest vocabulary word.A more discriminative and representative visual vocabulary is produced through adding constraint condition to nearest neighbor distance.Quantization error using this visual vocabulary is small.The proposed approach meets real-time constraints.Experiments based on datasets from dynamic environments and visually repetitive environments demonstrated that the largest recall rate increased by 5%and 4%respectively at 100%precision when compared with the state of the art.Information matrix used inside COP-SLAM is computed based on the number of image feature inliers.The obtained information matrix is used as a quantitative measure of visual odometry accuracy.However,it can not be a good measure of visual odometry accuracy when inlier points are mainly distributed on the specific region of the image,which affects the optimization results.A general approach to graph optimization algorithm is proposed based on inlier distribution.Inlier distribution is represented by the area of inlier set.Calculation of information matrix is not only based on the number of image feature inliers but also the area of inlier set,which improves the measure of visual odometry accuracy.The experimental results demonstrated that the method could effectively reduce the absolute trajectory error.