A Fast and Efficient FPGA-based Pose Estimation Solution for IoT Applications
Xiang Wang, Zikang Zhang, Yiting Wang, Chang Cai, Gengsheng Chen · 2022 IEEE 65th International Midwest Symposium on Circuits and Systems (MWSCAS) · 2022
CNN-based pose estimation can be widely used in various edge computing applications. However, its complex computations severely bring about a high processing latency and a high demand in computing resources. In this paper, we present an FPGA-based heterogeneous SoC design for real-time pose estimation. First, on top of SimpleBaseline, a new CNN model – LightPose is proposed for an efficient and lightweight processing of human’s pose estimation. By using MobileNetV2 as its backbone for feature extraction and using depth-wise separable transpose convolutions for up-sampling, LightPose has successfully obtained an over $50 \times$ reduction in both computation complexity and network size. Second, by using a quantization-aware training method with an additional shortcut optimization, an 8-bit quantization is conducted on LightPose with an only 0.5% drop in average precision. Finally, with specially designed computing engines and pipelined modules, we build the LightPose model on a Xilinx xc7k325t FPGA together with a RISC-V CPU for system management and external communications. Experimental results show that our new LightPose model and its hardware-accelerated solution have successfully reached an outstanding performance of 411.6 FPS in speed and 0.546 in average precision, surpassing the existing peer works in both processing rate and power efficiency.