DSOD: DSO in Dynamic Environments

Abstract

Recently, visual simultaneous localization and mapping (SLAM) has been widely used in robotics and autonomous vehicles. It performs well in static environments. However, real-world environments are often dynamic scenarios. Because it is difficult for SLAM to deal with moving objects such as pedestrians and moving cars, SLAM does not meet the actual needs of robots and autonomous vehicles in real-world scenarios. Visual odometry (VO) is a key component of SLAM systems. In this paper, to extend SLAM to dynamic scenarios, we propose a monocular VO based on direct sparse odometry (DSO) to solve the problems arising in a dynamic environment. The proposed method, called DSO-Dynamic (DSOD), combines a semantic segmentation network with a depth prediction network to provide prior depth and semantic information. Experiments were conducted on the KITTI and Cityscapes datasets, and the results show our method achieves good performance compared with the baseline algorithm, DSO.