Real-Time Multi-Drone Detection and Tracking for Pursuit-Evasion With Parameter Search
Jiaping Xiao, Jian Hui Chee, Mir Feroskhan
Abstract
Real-time multi-object detection and tracking are primarily required for intelligent multi-vehicle systems. This paper presents a whole life cycle multi-drone detection and tracking approach for collaborative drone pursuit-evasion operations, incorporating parameter search and edge acceleration techniques. Specifically, to address the single-class drone detection limitation of existing drone datasets, we first collect a new dataset “ICG-Drone” from various environments and then establish a performance benchmark with different models, such as YOLOv5, YOLOv8, and Swin Transformer. Based on the outstanding performance regarding accuracy, inference speed, etc., the selected YOLOv5s is further fine-tuned with a genetic algorithm, which achieves 14.8% / 3.6% improvement over 2 drone classes and 3 drone classes, respectively, in terms of mean average precision (mAP). Moreover, we develop an edge-accelerated detector and tracking system Drone-YOLOSORT focusing on “evader” and “pursuer” drones using TensorRT and deliver a ROS package for modular integration, which can be easily applied in a multi-drone system for recognizing friends and non-friends. Our system is able to reach about 24.3 FPS during inferencing, fulfilling the criteria of real-time drone detection at 20 FPS. The project is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/NTU-ICG/multidrone-detection-tracking.git</uri> .