Temporal and spatial 3D motion vector filtering method for robust visual odometry

Abstract

Visual odometry is a popular approach for estimating the trajectory of a moving vehicle from video input. For this purpose, most of the previous methods have depended on the RANSAC method; however, the RAN-SAC method cannot deal with a large independently moving object that takes up over 50% of the image ar-ea. This paper describes a visual odometry algorithm that deals with the nearly degenerated situation caused by a false motion vector generated by independently moving objects, falsely matched points and wrong depth information that often arise in visual odometry for outdoor service robots. To filter out these motion vectors, the spatiotemporal filter uses the current and previous motion vector’s length and direction. Experi-mental results demonstrate that the proposed method effectively rejects the motion vectors generated from large independently moving objects, thus avoiding false matching result. The results verify that the proposed method can estimate poses accurately.