Dyna-VO: A Semantic Visual Odometry in Dynamic Environment

In traditional visual odometry, most of them use static scenes as their assumptions and do not consider the moving objects in the actual environment, such as pedestrians and vehicles. These moving objects will cause poor tracking effects when the image features are matched, which directly decreases the accuracy of camera pose estimation. Aiming at this problem, A visual odometry method combining semantic segmentation information in a dynamic environment has been proposed (Dyna-VO). The system applies the optical flow pyramid method to motion consistency detection and builds a convolutional neural network to obtain the semantic contour information of the objects. The identified potential dynamic features are matched with semantic segmentation, and the features in the contour are eliminated as a whole to reduce the influence of dynamic objects. The experimental results in the public dataset KITTI and the campus environment show that Dyna-VO, compared with ORB-SLAM2, significantly improves the accuracy of feature matching under a challenging and highly dynamic environment, and the pose estimation is more accurate.