Research On Autonomous Navigation Of Indoor Mobile Robot Based On Depth Camera

With the development of artificial intelligence and sensor technology,mobile robots equipped with vision sensors continue to emerge.However,in the indoor situation,facing the complex environmental information,the mobile robot needs to face the following problems when using the depth camera for autonomous navigation: One is that the location and the grid map needed for path planner with depth camera;The second is how to improve the efficiency of global ptah planning by grid map;Third,how to avoid obstacles for pedestrians in a dynamic indoor environment.Therefore,in view of these problems,this paper carries out research from the following aspects:Firstly,this paper analyzes the four coordinate systems of visual correlation and the relationship between these systems.Take advantage of this relationship,it can build the Kinect v2 camera model.According to the camera model,robot can obtaine the three dimensional data of the environment by the joint alignment of the color image and the depth image.Last,the principle of the camera’s distortion correction is analyzed.It provides a theoretical basis for the positioning and navigation system of mobile robots.Secondly,the classic framework of visual ORB SLAM2 is studied,and the visual odometry,key frame selection strategy,loop closing and pose optimization algorithms involved in it are analyzed in detail.Aiming at the problem that the classical ORB SLAM2 framework only maintains sparse point cloud map but is not suitable for navigation,the framework of this algorithm is extended.Octree structure is used to complete the construction of real time Octomap map,which clearly describes the the state of idle or occupied for the three dimensional environment.In the oblique projection mode,it can provide mobile robot with a two-dimensional grid map containing certain height obstacle information,which improves the usability of ORB SLAM2 in mobile robot system.Thirdly,based on the two-dimensional grid map,the traditional A* algorithm has many redundant points and low search efficiency in global path searching.Therefore,in order to solve these problems,this paper studies the optimization strategy of JPS algorithm for A* algorithm under the jump point search theory.In order to further improve the speed of the path search,the traditional bidirectional search strategy is improved in this chapter.Using this strategy,this paper put forward a suggestion that is bidirectional synchronous JPS algorithm,it can improve the efficiency of JPS algorithm in complex grid map.Then,there is always the problem of path selection without considering the longer time sequence from the previous local planner.This paper designs a dynamic obstacle avoidance system that integrates the pedestrian trajectory information for pedestrians in an indoor environment.It mainly uses the Yolo v3-Deep Sort deep learning network to detect and track pedestrians within the range perceived by the Kinect v2 camera,and predict the position of the pedestrian in the next cycle when the pedestrian trajectory information is known.According to the obstacle avoidance strategy of the improved rolling window method in this paper,the predicted pedestrian information is selectively updated to the multi-layer cost map for the DWA algorithm to predict and improve the effect of dynamic obstacle avoidance.Finally,in the indoor environment,grid map construction and navigation performance test based on three wheel omnidirectional mobile robot.Through these experiments,the algorithm proposed in this paper is verified.According to the results,it can prove that the scheme proposed in this paper can run well in the actual environment,and achieve the desired goal.