Pixel-based object motion detection and tracking with a moving camera

Abstract

Object motion detection and tracking is a major research focus in visual autonomous navigation. In the case of a moving camera, pixel-based approaches to detecting motion are difficult as the algorithm must be able to distinguish between pixel motion caused by ego motion and pixel motion caused by object motion in the environment. In this study, we propose a low-computation pixel-based method to detect and track moving objects in scenes with a moving camera. Our method calculates deviations in pixel values over a stream of RGB images and thresholds the deviations to identify motion. Most importantly, the method accounts for ego motion by using depth information and camera poses to determine the best pixel locations to compare across images. We experiment with two different methods of calculating deviation: standard deviation and Gaussian distribution percentile. Our proposed method is evaluated on a dataset from the AirSim Safari Environment and shows the ability to detect and track only the moving objects in scenes.