Target Tracking via Deep Learning

To address the challenge of long-term tracking, through extended occlusions and significant appearance changes, we propose to continue developing DC-CAT, a Deep Convolutional neural network (CNN) based Confuser-Aware high value target (HVT) Tracker.The DC-CAT system combines a state-of-the-art CNN-based adaptive HVT tracker with a CNN-based pre-trained generic target detector, in a deep-feature-aided multi-target tracking (MTT) framework. The generic target proposals outputted by this second detector enable the tracker to (1) detect when the HVT is occluded and avoid updating its adaptive discriminative model during the occlusion, and (2) re-acquire the target afterward, even if the targets appearance has changed significantly. The multi-target tracker provides awareness of confusers, greatly reducing the likelihood that the tracker will be diverted to a confuser during the occlusion event or in dense traffic.Our team, consisting of Systems & Technology Research (STR) and Rochester Institute of Technology (RIT), leverages extensive experience in tracking algorithm development for real-time airborne surveillance applications, as well as our expertise in deep-learning-based object detection and tracking, to develop a robust tracking capability that is adaptable to a wide range of target, scene and sensor operating conditions.