Research And Implementation Of Visual Semantic SLAM Technology In Dynamic Environment

Simultaneous localization and mapping(SLAM)technology is highly significant in enabling smart devices to explore unknown environments.However,traditional SLAM technology relies on strong static assumptions and does not perform well in high dynamic environment.Existing visual SLAM solutions for dynamic environment leverage deep learning techniques to segment static and dynamic regions and remove features in dynamic areas.This approach can enhance the accuracy of visual SLAM in dynamic environments to some degree.Fewer constraints available due to dynamic points being culled,and there is a need for improving the robustness of the system.This paper is focused on visual SLAM technology,with a particular emphasis on addressing the SLAM problem in dynamic environments.The paper proposes an algorithm for pose solution,which is based on dynamic and static region segmentation,as well as a visual SLAM method based on multi-objective motion estimation.The pose solution algorithm employs the optical flow method to compensate the instance segmentation result of the deep learning model.This can help alleviate the issues that caused by the missed detection of deep learning method.Different from other visual SLAM solutions in dynamic environments,this paper does not require waiting for each frame to complete deep learning model inference.Instead,it integrates deep learning into the system in a non-blocking way.The dynamic and static areas of the scene are segmented based on observation semantic information and the relationship between map points and features.Camera pose constraints are constructed by combining the static feature reprojection residual and the relative position residual to solve the camera pose.Building on the foundation of dynamic and static area segmentation,this paper proposes a visual SLAM method based on multi-objective motion estimation.Instead of directly removing dynamic constraints,this method utilized dynamic object information to enhance the robustness of the camera pose solution.In this paper,the feature similarity between objects is calculated according to the matching of the object feature set,and the object position similarity is evaluated based on Kalman Filter,the object matching cost is calculated from the feature similarity and position similarity.And then the Hungarian algorithm is used to obtain the best matching of multiple objectives.To solve the camera pose,a joint optimization problem of multi-objective motion and camera pose is constructed,this paper refines both simultaneously.Moreover,this paper designs and implements a visual SLAM system for dynamic environments,providing users with functions such as simultaneous localization and mapping,evaluation of localization accuracy.The system’s core function is based on the visual SLAM method proposed in this paper.The experimental results show that the proposed method outperforms existing visual SLAM systems in terms of accuracy,real-time performance,and robustness when dealing with dynamic environment,and the fusion of multi-motion estimation makes the system more universally robust.