Research On Algorithm Of Dynamic Objects Tracking Based On Visual SLAM

Simultaneous Localization and Mapping(SLAM)is an environment perception technology.When a device is in an unknown environment,understanding the environment will be the primary task of the device.SLAM technology is one of the key technology for it to perceive surrounding information.Visual SLAM technology using cameras as sensors has always been a hot research topic.Traditional visual SLAM technology is based on the static assumption of the environment,assuming that the image features captured by the camera are always in a stationary and unchanging state,ignoring the disturbance caused by dynamic features.In order to endow the visual SLAM system with the perception ability of dynamic objects and improve the accuracy of SLAM in dynamic environments,this article proposes a visual SLAM-based dynamic object tracking algorithm.The algorithm is divided into three parts: data preprocessing,visual odometry,and factor graph optimization.The specific contents of the research are as follows:(1)In response to the issue that the original stereo images are difficult to process directly,relevant technologies from computer graphics are used for information preprocessing.This part mainly includes four parallel modules: depth generation,potential dynamic objects recognition,optical flow estimation,and feature point extraction.After processing through this part,the algorithm can finally extract four types of information needed for subsequent processing from the original stereo images: depth,semantics,optical flow,and feature points.(2)In response to the issue that traditional visual SLAM algorithms cannot estimate the position of dynamic objects,a visual odometry with dynamic object tracking function is designed.The algorithm first completes the data association of point-to-point and object-toobject based on optical flow information and pixel point ownership.Secondly,an improved dynamic object recognition method is used to complete the perception of dynamic features.Then,static features are used to solve the camera pose to eliminate the interference of unstable feature points on the positioning process.Finally,the motion between dynamic objects and the camera is decoupled,and the motion of the dynamic object is solved by a nonlinear optimization method using an implicit description of the motion transformation of the objects point.After processing through this part,the algorithm can improve the accuracy of SLAM in dynamic environments and track the motion of dynamic objects.(3)In response to the issue that traditional visual SLAM algorithms can only optimize two types of state variables,camera pose and map point,an improved graph optimization algorithm is proposed to jointly optimize camera pose,map point,and object motion.The factor graph model in this article mainly includes the static feature part and the dynamic feature part.The static feature part mainly optimizes the map points and camera poses.The dynamic feature part mainly optimizes the transformation pose of the dynamic object.By integrating all error terms and constructing an optimization function,the variables are optimized.After processing through this part,the algorithm can jointly optimize the three types of state variables of the camera pose,map point,and transformation pose of the object,and improve the accuracy of the optimization results.Experimental results show that the algorithm proposed in this article can complete camera positioning,map construction,and dynamic object tracking.In addition,the algorithm improves the accuracy of SLAM in dynamic environments and the effectiveness of dynamic object tracking.