A Power-efficient end-to-end Implementation of YOLOv8 Based on RISC-V

You Look Only Once (YOLO) is a highly popular object detection framework known for achieving an optimal balance between speed and accuracy. It has gained widespread adoption across diverse domains due to its exceptional performance. In this paper, we present an end-to-end runtime-configurable DNN accelerator based on the RISC-V architecture targeting at YOLOv8s model with Int8 precision. Additionally, neural network layer scheme optimization, piecewise linear function Approximation, data per-group quantization and Batch Normalization merging are introduced. The implemented model achieves an input resolution of 256 x 256 and exhibits excellent power efficiency with a theoretical throughput of 114 fps, which opens possibilities for real-time object detection in resource- constrained environments.Beyond its exemplary energy efficiency, the proposed architecture offers adaptability to different variants within the YOLO family, allowing for smooth integration with various YOLO model variations, enhancing its utility across diverse applications. The paper’s contributions extend to fields such as computer vision, robotics, smart surveillance systems, and autonomous vehicles, where real-time and accurate object detection is essential.