Small Moving Target MOT Tracking with GM-PHD Filter and Attention-Based CNN

Abstract

We present a multi-object tracking (MOT) approach to track small moving targets in satellite images. Our objects of interest span few pixels, do not present a defined texture, and are easily lost in cluttered environments. We propose a patch-based convolutional neural network (CNN) that focuses on specific regions to detect and discriminate nearby small objects. We use the object motion information to drive the patch selection and detect objects using a region-based CNN. In addition, we present a direct MOT data-association approach by using an improved Gaussian mixture-probability hypothesis density (GM-PHD) filter. The GM-PHD filter offers an efficient yet robust MOT formulation that takes into account clutter, misdetection, and target appearance and disappearance. We are able to detect and track blob-like moving objects and demonstrate an improvement over competing state-of-the-art tracking approaches.