Uav Detect, Track and Follow (DTF) of Non-Stationary Targets in Aerial Thermal Videos

Evolution of portable Graphics Processing Units (GPUs) have rejuvenated the research on real-time target tracking in aerial systems. These devices are highly optimized in terms of volume, size, power, weight and are effectively employed in UAV based applications. Tracking a target can be broadly divided into two sub-categories: (a) Surveillance and Tracking of the target from a distance and (b) Track and Follow towards the target(e.g. medical and e-commerce deliveries), where the UAV has to intercept the target as closely as possible. Surveillance applications are extensively studied and successfully implemented in real-time. However, present day target follow applications rely completely on the GPS data for intercepting the target. In this work, we propose a novel convolutional neural network (CNN) based target follow in GPS denied environments. The proposed application presents novel challenges in real-time with changes in aspect ratio, target orientation, size of the target in the Field of View (FOV) and track loss. This is further exacerbated in case of thermal systems due to limited information and low thermal contrast ratios between foreground and background. In this work, we address the problem of target follow specifically in thermal videos. We deployed the proposed solution on a portable NVIDIA GPU integrated on a custom developed UAV. The algorithm proposed is successfully tested and validated and the results obtained demonstrate the efficiency of the algorithm for aerial thermal datasets.