Edge-Optimized Real-Time Object Detection in AIoT Systems Using Quantized YOLOv8 and Deep SORT

Abstract

Autonomy and real-time monitoring are the backbone of today's smart infrastructure, which AIoT powers. Problems with detecting accuracy, tracking stability, and system feasibility on edge devices are common in existing systems. For accurate, low-latency multi-object identification on platforms with limited resources, this study suggests EdgeTrack-YOLOv8(You Only Look Once), an intelligent monitoring framework based on the artificial intelligence of things (AIoT) that combines Quantization-Aware Training (QAT), Edge-Aware Attention (EAA), and Deep SORT(Simple Online and Real-time Tracking) tracking. The first step in the process involves gathering and enhancing a 6,603-image surveillance dataset. Step two is training an enhanced YOLOv8 model using QAT and EAA to improve edge efficiency. Step three is applying Deep SORT to provide strong identity tracking even when faced with occlusion and varied lighting conditions. The results show that the NVIDIA Jetson Xavier NX is capable of 3458.4 mAP/J sustained detection efficiency, 89.6% mAP@0.5 (a 4.2% improvement over baseline YOLOv8), 32 FPS real-time inference, 18% latency reduction, 92.7% occlusion persistence, and 32 FPS(Frames Per Second)input lag. The results show that EdgeTrack-YOLOv8 can deploy AIoT in dynamic surveillance situations accurately, efficiently, and in a scalable manner.