UAS-based object tracking via Deep Learning

Unmanned Aerial System-based object tracking is a challenging new task in the computer vision community. In addition, existing benchmark and research focus on short sequences that are less than a minute long. In this work, we show the limitations of state-of-the art trackers in front of long-term aerial tracking. We propose a novel long-term, real-time, intelligent system for unmanned aerial system -based vehicle tracking utilizing deep learning techniques. We integrate a fast and accurate correlation filter with the expressiveness of a convolutional neural network embedding. A re-initialization policy based on a real-time anomaly detection on correlation map combined with a one-shot detector ensure that our system is impervious to drift and occlusion.