Hybrid Optimization of Target Detection on Embedded Platforms for Real Time Applications

Target detection has been widely used in fields such as intelligent security and autonomous driving. However, existing computationally heavy target detection algorithms based on deep learning can only work on GPU and CPU platforms, restricting the applications on edge devices with limited computational power. To address this issue, this paper proposes layer fusion and 16-bit fixed-point quantization on the YOLOv2-Tiny algorithm to reduce the computational complexity of target detection algorithms. Furthermore, the data transmission efficiency is optimized by using ping-pong butter and multi-channel methods. To reduce FPGA resource consumption, the neural network is split into convolution, accumulation, pooling, and address mapping modules. The proposed system has been successfully implemented on the Xilinx Zynq-XC7Z035 platform, using only 47% of BRAM resources and 18% of DSP resources.