Edge-based Object Detection using Optimized Tiny YOLO on Embedded Systems

Object detection at the edge has gained considerable attention for enabling real-time, low-latency, and privacy-preserving solutions by processing data locally on resource-constrained devices. This paper explores using Tiny YOLO, a lightweight variant of the YOLO architecture, for object detection on embedded systems. Tiny YOLO is specifically designed for edge devices to run efficiently on constrained devices by utilizing a reduced architecture with fewer parameters while maintaining good performance for real-time object detection. The study examines the deployment of optimized Tiny YOLO models on embedded systems, incorporating techniques like quantization, pruning, and clustering to reduce model size, enhance speed, and lower power consumption. Optimization methods show significant improvements, with quantization speeding up inference, pruning eliminating redundant parameters, and clustering enhancing accuracy. Specifically, the study compares the performance of Tiny YOLO under these optimization techniques, presenting results for both Pascal VOC and COCO datasets. The results demonstrate that optimized Tiny YOLO models are effective for real-time object detection on microcontrollers. These methods enable the efficient deployment of deep learning models for edge computing, without relying on cloud infrastructure.