Slam For Autonomouse Vehicles Based On Stereo Vision In The City Scene With Dynamic Objects

Simultaneous localization and mapping refers to the estimation of the position and attitude of a moving carrier in real time and the establishment of a map of the surrounding environment by using the data of sensor it carried in a strange environment.The technology is the basis for intelligent unmanned system to realize autonomous navigation,and has high research value.In the outdoor environment,stereo camera has become an important means to realize simultaneous localization and mapping because of its abundant information and low cost.However,due to the complexity of the actual situation,this field is still faced with some challenges,such as rapid movement,illumination changes and dynamic interference problems,so simultaneous localization and mapping for driverless car with stereo camera in city scene is studied in this paper.Firstly,the simultaneous localization and mapping method based on feature matching is studied,and the basic principle and system flow of thi s area are introduced.How to extract the stable feature points from the image is studied then in detail,so as to determine the correlation between the observation data when the environment illumination and the view angle of camera change,that is,to complete the matching between the pixels.We introduces how to calculate the position and attitude of the camera according to the coordinates of the matching points,and studies the filtering method of the key frames used to construct the map and how to detect the loop based on the bag of words.Secondly,a method to remove the dynamic feature points of the system based on semantic grouping is proposed to deal with the interference of moving objects to the eastimation of camera pose in city scenes.How to use computer vision technology to get semantic information from single frame image to group feature points is introduced.In the process of simultaneous localization and mapping,we use the geometric constraints between frames to calculate the reprojection error of feature points.According to the semantic information,the mean errors of the sample in each group is calculated to amplify the diffreence between in the static and dynamic difference,enables the system can not only reduce the interference of dynamic objects effectively,but also keep enough static feature points to locate at the same time.Nextly,aiming at the problem that the map established by the simultaneous localization and mapping system based on feature matching can not effectively reflect the environment structure and is difficult to meet the requirements of navigation and obstacle avoidance,a method is designed to construct the 3 D dense map through the stereo matching of key frames.We introduces the generative model of left and right images and use the model to estimate the dense disparities based on the disparities of feature points as a priori,and gives the post-processing process of the disparities map and the method to detect the outlier in 3D map according to the covisibility graph.Finally,the improvement of th simultaneous localization and mapping system is verified by qualitative and quantitative experiments.The test results on KITTI,the largest data set for algorithm evaluation in automatic driving scenarios show that our algorithm designed in this paper can effectively improve the positioning accuracy of the system,and can construct a 3D dense map that clearly reflects the detailed structure information of the environment.