Research On Visual SLAM Technology In Dynamic Environments

With the development of industrial technology and information technology,robots are widely used in different fields,and the human demand for intelligent medical service robots with autonomous navigation capabilities is increasing.The ability to navigate autonomously relies on the robot’s estimation of its own motion and knowledge of its surroundings,so simultaneous localization and mapping(SLAM)algorithms have become a key technology in the field of robotics.With the development of the computer vision field,cameras are becoming the mainstream sensors in SLAM systems.Many classical and effective visual SLAM solutions exist,but most algorithms are based on static world assumptions,which greatly limit the application of algorithm to dynamic and complex real-world scenarios.To address the robustness problems encountered by visual SLAM algorithms in dynamic environments and dynamic object tracking problems,the main work of this paper is as follows:(1)Proposed visual odometry based on segmentation and optical flow to improve the robustness of bit-pose estimation.Most current SLAM algorithms assume that the camera is in a static scene,and when the dynamic part of the scene increases,the accuracy of camera localization decreases.In this paper,we use an instance segmentation network to detect a priori dynamic objects in images,and use dense optical flow to pass instance segmentation information and complete static feature point matching between two adjacent frames to achieve more accurate feature matching in repetitive texture scenes.This visual odometry has no loss of accuracy and faster operation than previous methods based on geometric information and target detection information.(2)Proposed adaptive coupling strategy to optimize camera and dynamic object poses.Simultaneous localization and object tracking are the extended problems of SLAM in dynamic scenes.Understanding the motion state of dynamic objects in the scene can enhance the robot’s understanding of the scene and improve the problem of SLAM algorithm failure in highly dynamic scenes.This paper adaptively selects the coupling method according to the dynamic degree of the scene:using a loosely coupled approach to optimize the system when the dynamic degree of the scene is small and ensuring the accuracy of the system through local optimization threads;using a tightly coupled approach when the dynamic degree of the scene is high and adding constraints to the camera self-motion estimation by using the positional transformation of the feature points on the dynamic objects in the scene.The system operation speed is accelerated while ensuring the system accuracy.(3)Combining front-end visual odometry and back-end adaptive coupling optimization strategies to build a visual SLAM system for dynamic environments.The system uses a point set-based object motion model to convert the complex dynamic object localization problem into a dynamic point pose estimation problem.The system operates fast by estimating the camera pose to construct a peripheral map while localizing and estimating the pose of a priori dynamic objects in the scene.In order to verify the performance of the system,the system is tested on the KITTI Tracking dataset,and the experimental results show that the system has good performance in camera pose estimation and object motion estimation in dynamic scenes.