Moving Object Segmentation by Combing Deep Learning and Geometric Constraints for Autonomous Driving

Moving object segmentation is an important and challenging task in the field of autonomous driving. This paper presents a novel and effective method that combines deep learning and geometric constraints for moving object segmentation. First, potential moving objects in the traffic scene are segmented using SOLOV2. Next, the residual optical flows (ROFs) of feature points on the potential moving objects are calculated. Lastly, a strategy considering randomness of ROF angles and ROF magnitude thresholding is developed to estimate the motion status of these potential moving objects. The effectiveness of the proposed method is evaluated on the KITTI and KITTI MOD datasets, and the experimental results demonstrate that the proposed method can achieve higher precision and recall compared to the existing algorithms.