Visual SLAM Algorithm Based On Improved YOLO And Optical Flow Constraints

Simultaneous Localization and Mapping(SLAM)refers to a mobile robot synchronously determining its own location and incrementally constructing an environmental map through the information obtained by the onboard sensors without any prior information of the environment.Visual sensor(such as camera)has low cost,small size and abundant information.Therefore,vision-based SLAM system has become a hot research topic in the field of robotics.Early visual SLAM systems are based on the assumption that the environment is static.However,with the complexity of the robot working environment,this ideal assumption is no longer applicable,and researchers have turned to the research of visual SLAM algorithm in dynamic environment.When the robot runs in a dynamic environment,it will encounter a large number of dynamic targets,such as people or pets walking in the room.If the dynamic feature points on the dynamic target are not filtered in the feature matching process,pose estimation errors will occur,resulting in a decrease in the positioning accuracy of the robot.Therefore,this paper proposes a visual SLAM algorithm based on improved YOLO(You Only Look Once)and optical flow constraints,aiming at the problems of low positioning accuracy and poor mapping effect of visual SLAM systems in dynamic environments.The main work contents are as follows:(1)In order to effectively detect dynamic targets in the scene,a dynamic target detection algorithm based on improved YOLO-v3 and optical flow constraint is proposed.YOLO-v3 was first used to detect potential dynamic targets in the scene.Since the small number of feature points on the small target will not seriously affect the system accuracy,the original YOLO-v3 network structure is optimized,the feature extraction layer for detecting large and medium targets in the network is retained,and the network detection category is modified to make it only the common dynamic targets in life are detected,which saves the detection time and improves the detection efficiency.In order to further determine which of the detected targets belong to dynamic feature points,the dynamic probability of each feature point is calculated based on the optical flow constraint,and the dynamic feature points and static feature points are classified according to the dynamic probability.Finally,all non-conforming feature points were filtered based on geometric constraints.The experimental results show that the proposed algorithm can effectively detect dynamic targets in the scene.(2)Based on the ORB-SLAM2 framework,the above dynamic target detection algorithm is added to the original framework as a new parallel thread,and the dynamic feature point filtering algorithm is added to the tracking thread to filter the detected dynamic feature points,and only static feature points are used for feature matching to ensure the tracking accuracy.At the same time,an adaptive algorithm for increasing the number of feature points is added,which increases the number of feature points according to the area occupied by the dynamic region to ensure the stability of tracking.(3)The comparative experiments are carried out on public datasets,which are compared with the classic static visual SLAM and dynamic visual SLAM respectively,and the experimental results are evaluated from the two criteria of absolute trajectory error and relative rotation error.The experimental results show that the visual SLAM algorithm proposed in this paper can effectively filter out the dynamic feature points on the dynamic target,reduce the impact of the dynamic target on the system,greatly reduce the error of pose estimation,and improve the positioning accuracy.