A robust vehicle tracking in low-altitude UAV videos

In this study, we concentrate on solving variations in scale, aspect ratio, rotation, visual model and target motion problems for vehicle tracking in low-altitude UAV videos. The contributions of this work are threefold: 1. By introducing a particle rescaling mechanism where each particle is resized with different aspect ratios, tracking under scale and aspect ratio variations is improved. 2. By fully integrating a particle filter with a convolutional neural network, a new structure, which acts as an auxiliary particle filter, is developed. This new structure improves the estimation of the states, namely the location and the velocity of the target, and the dimensions of the bounding boxes, thus enables tracking under fast motion. 3. By introducing a unified multi-part vehicle tracking framework, robust tracking is achieved against scale change, aspect ratio, visual model variations and sudden rotations. The processing of multiple parts, independently, improves the tracking under sudden aspect ratio and rotation changes compared to tracking the vehicle as a whole. In this study, without loss of generality, the number of independent parts is taken as two and the proposed method is tested for image sequences from UAV dataset with various visual problems. The comparisons with the state-of-the-art trackers show that the proposed method achieves good precision and success scores, and outperforms most of the state-of-the-art trackers.