Real-Time Object Detection Using Edge Quantized You Only Look Once Version 5-Nano Framework

Currently, object detection on edge devices is a crucial component of smart surveillance, autonomous systems, and Internet of Things (IoT)-based real-time monitoring. Thus, You Only Look Once (YOLO) models have emerged as popular solutions owing to their speed and accuracy. However, the existing Field Programmable Gate Array (FPGA)-based YOLO focuses on single-class tasks and lacks scalability for multiclass detection in dynamic environments. Hence, this research proposes an Edge Quantized YOLOv5-Nano (EQ-YOLOv5) framework for scalable object detection and optimized edge deployment. Initially, the input data were collected from the COCO 2017 dataset, which contained over 80 object classes with labeled images. Then, the data are preprocessed with bilinear interpolation, min-max normalization, and bounding box normalization to resize images, normalize pixel values, and convert annotations to YOLO format. Subsequently, the model was trained with Quantization-Aware Training (QAT) to maintain accuracy under the INT8 constraints. Subsequently, the model was converted to a frozen format, quantized, and compiled for hardware compatibility. Finally, it was deployed using dynamic class filtering for real-time object detection on edge platforms. The proposed EQ-YOLOv5 achieved better results in terms of precision (0.753) and recall (0.76) than existing YOLO-based pruning and quantization models.