Loop Closure Detection Of Sparse Direct Odometry Based On Sparse Direct Method

In recent years,the development and application of mobile robot technology is affecting people's lives and changing people's working methods.Nowadays,with the rapid development of modern high-tech technology and the continuous improvement of intelligent manufacturing level,the development of mobile robot technology is also moving towards the direction of intelligence and autonomy.One of the key technologies of mobile robot--Simultaneous Localization And Mapping(SLAM)is becoming a research hotspot.This technology can make robot move autonomously in unfamiliar environment.With the rise of the 5G and assisted driving,a large number of researchers focus on slam,so slam has important theoretical significance.One of the visual SLAM methods,feature method,also known as indirect method,can use the image feature information to identify,match,locate and build a map,which is one of the common methods of visual SLAM.However,visual SLAM uses camera as sensor,which is easily affected by illumination,brightness,angle and other factors.Feature method can not completely solve this problem.Visual direct method directly uses image pixels,gradients and so on,which can minimize the impact of camera itself,but it can not reuse image information,and can not use loop closure detection and back-end optimization of global trajectory,which is easy to cause problems long term cumulative drift affects the positioning accuracy.Based on the sparse direct method,this paper proposes a hybrid gradient-feature point as a new feature of visual odometry,and uses the new feature to use loop closure detection to reduce the trajectory error.The multi sub-map system is used to manage multiple sub maps,and a new sub-map is opened when the tracking is lost,so as to reduce the relocation error and improve the robustness of the algorithm.The main contents of this paper are as follows(1)A hybrid gradient-feature point is proposed,and ORB algorithm is used to extract ORB features from the image.The image is divided into several grids,and gradient-features are extracted from each grid by using image gradient.In order to make full use of the image information to improve the positioning accuracy of the algorithm,the mixed gradient point is used to estimate the camera motion,local mapping and loop closure detection.(2)Build a multi sub-map system to manage the sub-map,extract and match features in the current sub-map.After the camera relocation fails,build a new sub-map to continue tracking the camera.The multi sub-map system detects that the two sub-maps have overlapping parts,and merges the two sub-maps into a new sub-map to replace the original two sub-maps.The experimental results show that the multi sub-map system is effective and can improve the trajectory accuracy and algorithm robustness.(3)By improving the loop closure detection algorithm,a loop closure detection algorithm for multi sub-maps system is constructed.In order to improve the reusability of image scene and the detection efficiency of loop closure,the mixed gradient feature points are used to determine the image keyframe and detect the loop closure in the multi sub-maps system,so as to reduce the global trajectory drift and improve the positioning accuracy.Experimental analysis shows that this paper can improve the positioning accuracy and real-time operation of the system.