基于Zynq平台的图像目标检测系统

Image Target Detection System Based on Zynq Platform

摘要:

针对图像目标检测任务中采用的深度学习网络复杂的计算和规模庞大的计算参数,导致基于ARM架构的嵌入式系统上,目标检测任务存在着高延时和处理速度慢的问题,文章提出并设计实现了一种新型完整嵌入式道路车辆检测方案。该方案在基于YOLOv3-Tiny的特征提取网络中采用结构重参数化的方法提升模型检测精度,并通过Vitis-AI在Zynq嵌入式平台上部署DPUCZDX8G架构的加速核对卷积神经网络的并行加速,最后将改进的YOLOv3-Tiny网络模型经过量化、编译,以动态链接库的方式部署。实验结果表明,在VOC2007上测试最终实现均值平均精度(MAP)为0.597,实时处理速度为27.7FPS,同时帧率功耗比为1.49,适合边缘计算设备的低功耗要求。

Abstract:

Due to the complex computation of the deep learning network and the huge computational parameters used in the road vehicle target detection, the problem of high delay and slow processing speed exists in the target detection task on the embedded system based on ARM architecture. Aiming at the above problems, a complete embedded road vehicle target detection solution was designed and implemented in this paper. The structural re-parameterization is used in the YOLOv3-Tiny-based feature extraction network to improve the model detection accuracy, and the parallel acceleration of the convolutional neural network was deployed by Vitis-AI on the Zynq embedded platform with the DPUCZDX8G architecture acceleration core, and finally the improved YOLOv3-Tiny network model was quantified, compiled and deployed as a dynamically linked library. The experimental results show that the MAP of VOC2007 is 0.597, and the real-time processing speed is 27.7FPS. At the same time, the frame rate power consumption ratio is 1.49, which is suitable for the low power consumption requirements of edge computing devices.