Research On Visual SLAM System Based On Object Detection In Indoor Dynamic Environment

Intelligent mobile robots play an important role in industry,agriculture,military and other fields in the last few years.SLAM(Simultaneous Localization and Mapping)draws much attention from scholars because it enables robots to perceive the environment and acquire their location.Camera-based Visual SLAM has become a research hotspot due to its advantages of low cost and rich information.As the core technology of mobile robot,Visual SLAM technology in static environment has achieved fruitful success,and can provide good localization and take good mapping effect.However,when there is interference from dynamic objects in the environment,wrong association will be generated,resulting in the degradation of SLAM positioning accuracy and the existence of double shadow in map construction,influencing the overall system’s stability.The map established can’t be used by the robot to perform advanced tasks.At present,the real-time performance of SLAM system in dynamic environment is poor.Aiming at the above problems,the main research content of this paper is robot Localization and mapping(SLAM)based on object detection in indoor dynamic environment by using RGB-D camera.In dynamic environment,lightweight object detection network YOLOX-S is used in this paper to ensure the real-time performance of the system,while realizing the detection of prior moving objects.The detection results are combined with the multi-view geometry method to eliminate the interference of moving objects,and finally the static three-dimensional map is established.This paper’s main research work is summarized as below:1.Aiming at the problem that there are dynamic objects in the environment and the real-time performance of visual SLAM system in the current dynamic environment is poor,this paper introduces an object detection network with good real-time performance.The results of object detection are applied to SLAM.2.A visual odometry suitable for dynamic environment is proposed.The visual odometry combines object detection and multi-view geometry to filter dynamic points.Object detection network YOLOX-S can realize fast detection,but it has two disadvantages: first,it can only provide bounding boxes.If all feature points in the frame are removed,the number of feature points will be too small,which is detrimental to the system’s precision and stability.The second is the inability to deal with moving objects that are assumed to be stationary.In this paper,a dynamic point screening strategy is proposed in the visual odometry section,which reduces the influence of dynamic factors in the environment while preserving static information in the environment as much as possible.3.Investigate the dense point cloud map and Octomap generation methods in the dynamic environment.At present,most map is built on sparse point cloud,which result in a lot of information loss,and its application is very limited.In this paper,the impact of dynamic objects is removed by object detection,and dense point cloud map and Octomap are generated on the basis of key frames.4.Conducted localization and mapping experiments on TUM public dataset.According to the experimental results,the proposed algorithm can significantly improve the accuracy and mapping effect of SLAM in dynamic environments.At the same time,the system has good real-time performance due to the use of lightweight object detection network.