【摘要】 随着神经网络技术的发展，深度学习在目标检测领域取得了非常好的性能和应用效果，但深度学习算法的边缘部署一直是一个挑战。为了克服这一挑战，设计了一种基于YOLOv3-Tiny网络的实时目标检测硬件加速器，旨在实现高效准确的目标识别和跟踪，并满足边缘计算对性能和功耗的要求。为了减少FPGA在浮点运算上的开销，将神经网络的浮点数量化成了定点数进行计算；同时使用了DSP乘法优化、深度可配置的行缓冲机制，流水与并行划分等一系列的策略来提高性能。在口罩识别与运动跟踪以及COCO数据集的基准测试中，该文的加速器功能得以实现，有效算力为107.5 GOPs，帧率可达27.6 fps。更多还原

【Abstract】 The evolution of neural network technology has led to significant advancements in the field of object detection through deep learning,yielding impressive performance and application outcomes.However,deploying deep learning algorithms at the edge has remained a challenge. To address this challenge,this paper presents the design of a real-time object detection hardware accelerator based on the YOLOv3-Tiny network. The aim is to achieve efficient and accurate object recognition and tracking while meeting the performance and power consumption requirements of edge computing.To minimize the FPGA’s overhead in floating-point computations,this paper quantizes the neural network’s floating-point numbers into fixed-point representations for computation. Additionally,a series of strategies such as DSP multiplication optimization,large-scale feature map tiling,pipelining,and parallel partitioning are employed to enhance performance. In the context of mask recognition,motion tracking,and benchmark testing using the COCO dataset,the proposed accelerator demonstrates correct functionality and effective computational power of 107.5 GOPs,achieving a frame rate of up to 27.6 fps.更多还原