Priority Scheduling Using Recurrent Quadrant Search for Handling Priority and ‘Pop-Up’ Targets in Aerial Videos

Target detection and tracking is a well established application in the radar (RF) domain. However, the same operating principles may not directly extend to image-based tracking systems. In the optical domain, the problem becomes challenging due to the lack of an active transmit signal (as available in radar processing), which actively maps the environment for target detection and track initiation. Passive imaging systems often require a separate assisted intervention in the optical processing channel for real-time target detection and track initialization. This limitation has been successfully addressed with the development of automated tracking algorithms using deep learning architectures like Convolutional Neural Networks (CNNs). CNNs are extensively employed in generic object classification and localization applications for single and multi-target scenarios. In multi-target scenarios, the deep architecture framework detects all the objects in the frame without any priority designation (eg. YOLO, VGG19 etc.). However, in multi-target scenarios where, handling selective targets is critical (eg. tracking an ambulance in traffic) or prioritized tracking is needed of a ‘Pop-Up’ target, that is an object entering the field-of-view (FOV)) (i.e. in autonomous driving systems, e-commerce deliveries etc.), the generic networks are not sufficient. In this work, we aim at addressing this specific problem of priority/ selective target tracking in aerial videos. We propose a novel CNN termed ‘PriorityNet’ with embedded priority information and a recurrent quadrant search algorithm for prioritized tracking and false alarm rejection in multi-target scenarios. The CNN and the associated algorithm is successfully deployed on a portable NVIDIA Jetson-Nano board (a small form factor GPU) integrated on-board a custom built UAV. The complete system is tested on multiple open-source and custom multi-modal data sets (RGB and Thermal). The results obtained demonstrate the efficacy of the proposed method for priority target tracking applications.