Research On Visual SLAM Based On Semantic Information In Indoor Dynamic Environment

In recent years,with the vigorous development of computer related fields,intelligent mobile robots have gradually entered daily life.In an unknown environment,mobile robots build sparse feature maps and dense maps based on visual Slam algorithm to complete the positioning and navigation task.However,the traditional visual SLAM algorithm is based on the assumption that the scene is static and less considers the dynamic objects in the actual scene,which reduces the accuracy of positioning and mapping.Moreover,the robot cannot obtain semantic information from sparse feature map and dense map,so it is difficult to perform high-level semantic tasks.In the indoor dynamic environment,in order to improve the perception level of the environment and the accuracy of map construction,a visual SLAM algorithm based on semantic information is proposed in this paper.The main research contents of this paper include:(1)Basic research on visual SLAM algorithm.The basic model and principle of visual odometer in visual SLAM algorithm are deeply analyzed.Aiming at the influence of dynamic objects on visual slam,a visual SLAM algorithm based on semantic information is proposed.Based on orb-slam2 algorithm,dynamic region detection and semantic segmentation modules are added to remove the influence of dynamic targets on pose estimation,and the indoor map containing semantic information is constructed by using the semantic information of the image.(2)Research on visual odometer based on dynamic area detection in indoor dynamic environment.Aiming at the problem that dynamic objects reduce the accuracy of pose estimation in visual odometer,a visual odometer based on dynamic region detection is designed in this paper.In this method,the dynamic region detection module is added to the visual odometer,and the dynamic region of the image is detected by combining the homography matrix and dense optical flow,so as to improve the accuracy of pose estimation in the dynamic environment.At the same time,the trigonometric function principle is used to avoid complex operations and speed up the extraction of orb feature points.In the dynamic region detection,the dense optical flow method is used to calculate the optical flow field information of the image,and combined with the homography matrix to reduce the optical flow change caused by the camera motion.The moving region in the image is determined by the dynamic optical flow threshold,the feature points in the dynamic region are eliminated,and the static feature points are used to complete the pose estimation.(3)Research on indoor map construction algorithm based on semantic information.In this paper,Bise Net V2 network is used to segment the image collected by the sensor and obtain the semantic information in the image.Aiming at the over segmentation problem when the semantic information of two-dimensional image is mapped to three-dimensional space,this paper proposes a region growth algorithm combined with depth information to optimize the three-dimensional segmentation and improve the accuracy of object semantic annotation.Aiming at the influence of dynamic target on map building,the dynamic target is determined by combining semantic information with dynamic area,and the map points constructed by moving objects are eliminated to build indoor map.Finally,the algorithm proposed in this paper and the orb-slam2 and ds-slam algorithms are evaluated and analyzed by using the open data set of tum and the actual dynamic scene.The experimental results of tum public data set show that the trajectory accuracy of this algorithm is better than orb-slam2 and ds-slam algorithms in dynamic scene,which improves the positioning accuracy in dynamic environment and the construction effect of static semantic map.Experimental results in real dynamic scenes verify the feasibility and robustness of the proposed algorithm.