THE FIRST PLACE SOLUTION FOR ICIP2023 CHALLENGE INFRARED IMAGING-BASED DRONE DETECTION AND TRACKING IN DISTORTED SURVEILLANCE VIDEOS

This work proposes a solution for detection and tracking of drones in distorted Infrared (IR) images. Our proposed solution combines stats-of-the-art object detection and tracking techniques with distortion classification technique to achieve robust performance in highly distorted videos. There are two main steps in our proposed solution. First, an online drones detection using state-of-the-art YOLOv8 object detection framework is combined with Distortion Classification network. Second, we use the detection proposals from YOLOv8 for temporal association to track the drones using popular state-of-the-art trackers namely, ByteTrack and Strong-SORT. Our solution was evaluated using the ICIP-2023 challenge dataset for drones detection and tracking. The ICIP-2023 challenge dataset consists of IR images and videos with challenging attributes likes different types of distortions, varying size of drones and abrupt drone motion etc. with complex scenes. Our solution works in all the challenging scenario captured in the dataset and our submission achieved the best performance among all entries in the challenge, setting a benchmark for the ICIP-2023 challenge dataset with detection Accuracy of 95% and mAP of 98% for drones and Tracking HOTA of 0.68 and IDF1 0.77 with real-time operation at 35FPS.