Motion segmentation in Moving Camera Videos using Velocity Guided Optical Flow Normalization

An obstacle avoidance system is a key feature of a robot navigation system. A capable avoidance system should consider obstacles movements into account. This study proposes a new approach for detecting motion from a video captured by a moving camera, a similar scenario happens in a robot use case. The process starts from acquiring two successive video frames and computes its optical flow image using gmflownet. Then, the semantic segmentation mask, as well as the estimated depth map, are also generated. After that, camera velocity is estimated based on the optical flow of points belonging to static objects. Next, the velocity information is used for optical flow normalization. The normalized optical flow is then fed to a DeeplabV3 network to obtain a motion mask. Finally, the motion and semantic mask are fused in the postprocessing stage to obtain the final mask. Experiments on video data indicate that the performance of the proposed method exceeds that of the standard one indicated by the average precision, recall, and IoU of non-zero results of 0.814, 0.719, and 0.6222, respectively.