Edge-Optimized Deep Learning Framework for Dynamic Unmanned Vehicles Detection and Tracking

Abstract

The increasing prevalence of drones alongside security threats related to these in various sectors, including military, surveillance, and delivery, has underscored the urgent need for advanced drone detection and tracking systems (DDTS) capable of efficiently operating in diverse environmental conditions. This research addresses the limitations of existing object detection models by developing a specialized novel dataset in LWIR (Long-Wave Infrared) and SWIR (Short-Wave Infrared) spectrum which was processed through various data augmentation techniques to improve drone detection and tracking capabilities under varying illumination conditions. To improve computational efficiency and detection accuracy, several variants of the YOLOv10 model including n,s,m,b,l, and x are optimized using advanced techniques like Ghost Convolutions, Pruning, TensorRT, and ONNX configurations for CPU and GPU of advanced computer and edge-devices including Jetson Orin Nano and Jetson Orin AGX. The comparative evaluations of various tracking algorithms are conducted alongside the optimized YOLO models, revealing that Botsort is the most effective algorithm for real time drone tracking due to its superior performance in dynamic environments. The optimized YOLO models across various hardware configurations were compared based on performance metrics like inference speed, frames per second (FPS), mean average precision, recall, F1 score, and accuracy. The Ghost Convolutions incorporated YOLO models outperformed their standard models with much fewer parameters making them ideal for real time applications. The findings of this research not only contribute to the existing body of knowledge in the field of computer vision and object tracking but also provide practical guidelines for implementing effective drone detection and tracking systems (DDTS) to improve security and public safety.