An Ensemble of Spatial Clustering and Temporal Error Profile Based Dynamic Point Removal for visual Odometry

Abstract

Visual odometry in the field of computer vision and robotics is a well-known approach with which the position and orientation of an agent can be obtained using only images from a camera or multiple of them. In most traditional point feature-based visual odometry, one important assumption and also an ideal condition is that the scene remains static. In environments with dynamic objects, this assumption can lead to erroneous pose estimations. Though most modern systems that use visual odometry are equipped with approaches which reduce the effects of dynamic objects, having a module that can help the system to have a pre-understanding on the behaviour of features can better reduce errors. This work proposes complementary approaches suitable for different types of situations that filter out dynamic points as an ensemble: clustering triangulated points intended for reducing the effects of moving objects, analysing the reprojection errors and checking the consistency of points in three-dimensional space. The techniques employ a stereo camera and filtered points are passed through perspective-n-point for pose estimation. The approaches are tested on publicly available TUM dataset, as well as self-collected datasets with sequences containing dynamic objects in the scene. Results confirm that filtering the points before using them for pose estimation reduces errors in the trajectory.