Research On Visual Odometry Based On RGB-D Multi-mode Semantic Segmentation

In the strange environment of indoor,the mobile robots must complete the autonomous positioning before autonomous navigation.The traditional positioning methods such as wheel odometry and laser odometry,which have the problems of large measurement error and high cost.From the perspective of accuracy and cost performance,the use of RGB-D cameras as visual odometry has great potential for development,because of riching geometric depth information.But visual positioning is susceptible to environmental interference.In this paper,a visual odometry based on RGB-D multi-modal semantic segmentation is designed to improve the camera's visual positioning ability in dynamic environment.The main work of thesis is as follows:1.Aiming at the problem that RGB single-modal semantic segmentation model is used to remove dynamic interference of person,and not fully utilize RGB-D camera information,an improved RGB-D multi-modal semantic segmentation network model is designed.The network model adopting the cross-entropy and Dice coefficient double supervised function,which can simultaneously utilize the texture information of the RGB image and the edge information of the depth image.The experimental results show that the network model effectively suppresses the influence of the imbalance of the number of pixels in different types of images,obtaining a more accurate segmentation effect.2.Aiming at the influence of visual odometry positioning accuracy on dynamic object interference,this paper designs a visual odometer that uses RGB-D multi-modal semantic segmentation to remove dynamic objects.The semantic segmentation module is integrated into the visual odometry of the ORB-SLAM2 system,tested on the TUM standard data set and the collected laboratory dynamic scene.The experimental results show that the visual odometry designed in this paper has good robustness in the dynamic environment of pedestrians walking.3.Aiming at the problem of low efficiency of visual odometry using RANSAC algorithm to eliminate mismatches,the paper based on the existing algorithm adds matching points to satisfy affine transformation as pre-check constraints,reduceing invalid operations for homography matrix.Compared with the traditional RANSAC matching purification,the experimental results show that the average time saving about 35%,improveing the speed of visual odometry calculated.