Object detection has been a significant challenge in machine vision systems from the past to the present. Various hardware-based accelerators have been utilized to enhance speed efficiency. The primary objective of most of these accelerators is to minimize idle states in DSP blocks. In this paper, a new architecture based on Convolutional Processing Elements (CPEs) is proposed, wherein weights are stored, circularly shifted in an internal CPE buffer and used to generate output feature maps. In this way, the idle states of DSPs are reduced by increasing data reuse in CPEs and decreasing external memory accesses. The number of CPEs used to accelerate a CNN depends on the required speed and available hardware resources; configurations of 16, 32, 64, 128, and 256 CPEs can be utilized to accelerate a desired Convolutional Neural Network (CNN). To demonstrate the effectiveness of the proposed architecture, it is applied to the YOLOv3-Tiny object detection CNN. Experimental results show that our proposed architecture with 128 CPE cores can operate at 62.8 frames per second on an FPGA Xilinx XCKU060 with a working frequency of 200 MHz, using 16-bit fixed-point representation. This approach results in only a 1% drop in mAP while utilizing 43.2K LUTs, 94.4K FFs, 26.73 Mbits of RAM, and 1364 DSPs. Furthermore, the number of external memory chips is reduced by 67% compared to the state-of-the-art systems.
