Research On Visual Location Method Based On Semantic Information Fusion In Dynamic Environment

Localization is one of the key technologies for autonomous navigation,accurate behavior and safe application of unmanned intelligence system,and it is also the basis for the agent to make correct decision and path choice according to the task.Among various localization technologies,visual SLAM has many advantages,such as strong autonomy,low cost and rich perceptual information.Most existing visual SLAM systems are based on the assumption of static environment,meaning that using static scene information to estimate its own motion.Dynamic objects in the scene will destroy the feature correlation of the system,thus reducing the accuracy of motion estimation and the robustness of the algorithm.In some practical scenes such as automatic driving,it is an important task to locate the dynamic objects in the scene while estimating the camera's motion.Aiming at these problems,we put forward a kind of semantic information fusion visual localization method in dynamic environment.This method can obtain the pixel-level semantic information of dynamic objects in the scene.On this basis,a direct method of visual odometry is constructed to estimate camera motion based on the fusion of image semantic information.Through the factor graph model to jointly optimize the camera pose and dynamic object pose to realize the simultaneous localization of the camera and dynamic objects in the scene.Specific research contents and contributions are as follows:(1)The dynamic object segmentation method based on convolutional neural network is proposed.In this paper,we use the semantic segmentation technology in deep learning to identify potential dynamic objects in the scene as prior information,and use the motion consistency detection method to determine the motion probability of the potential dynamic objects as observation information.These two kinds of information are fused in a Bayesian framework to realize the accurate recognition of the semantic information of dynamic objects in a complex environment.The method can accurately obtain pixel-level semantic information of a real dynamic area in a complex environment,and provides a basis for high-precision visual localization in a dynamic environment.(2)The visual location algorithm combining the semantic information of dynamic objects is presented.Aiming at the problem of low accuracy and poor robustness based on direct method visual odometry in dynamic scenes,the direct odometry integrating the semantic information of dynamic objects is proposed.The pixel-level dynamic object semantic information is combined to realize the extraction of candidate points in the static region.Pyramid pose tracking model and sliding window optimization model based on the prior information of dynamic objects can achieve high-precision localization of the camera in a dynamic environment.Experiments show that the localization accuracy of the proposed method is 71%?86% higher than that of the traditional direct method,and the three-dimensional point cloud image constructed after removing the dynamic target is clearer and more detailed.(3)A dynamic object localization model for visual scenes is constructed.The dense depth of the scene is estimated by using stereo images,and the real scale of the scene is recovered.The dense optical flow of the scene is estimated based on deep learning.Aiming at the motion solution failure caused by occlusion and excessive motion,the pose calculation equations of cameras and dynamic objects are constructed,and optical flow constraints are introduced on the basis of reprojection errors to improve the robustness of the system.Observation and cumulative errors in the system will reduce the accuracy of the pose calculation.The factor graph model is used to establish a unified cost function to jointly optimize the spatial point observation,camera and dynamic object pose.The camera motion estimation of the proposed method is basically equivalent to the existing systems.For dynamic objects in the scene,the positioning accuracy of the method in this paper is better than that of the existing methods.