基于MPSoC的轻量化汽车检测系统及硬件加速平台设计与优化

Design and optimization of a lightweight automotive detection system and hardware acceleration platform based on MPSoC

针对车辆分类检测在精度和实时性方面存在的挑战,提出了一项改进方案,以优化YOLOv5s模型,旨在实现轻量化的汽车检测。通过在MPSoC硬件架构的现场可编程门阵列(FPGA)上设计系统,成功打造了一个具备高精度、快速检测和低能耗的解决方案。为了使得模型更适合嵌入式设备部署,采用了MobileNetv3 Small替代YOLOv5s的主干网络,并引入卷积块注意力模块(CBAM)注意力机制和Inner-IoU Loss优化方法,使模型在轻量化的同时提升了检测精度和速度。改进后的模型相较于原始Yolov5s模型,平均精度均值(mAP)提升了14.8%,参数量减少了49.7%,模型体积减小了40.7%,计算量减少了48.9%,在NVIDIA 3060上,改进后的检测速度提升了48.8%,达到了82 fps。此外,还利用FPGA对YOLOv5s进行了硬件加速。经过优化的系统达到了45 fps的检测帧率,并保持了较高的精度和速度,这一系统易于部署,适用于智能交通系统,满足其高效实时监测的需求。

In response to the challenges regarding accuracy and real-time performance in vehicle classification detection,this study proposes an improved lightweight model for vehicle detection based on YOLOv5s.The objective to achieve a solution that balances high detection accuracy,swift detection,and low power consumption through a system designed on an FPGA within an MPSoC hardware architecture.In order to make the model more suitable for embedded device deployment,This research replaces the backbone network of YOLOv5s with MobileNetv3 Small and incorporates CBAM attention mechanism and Inner-IoU Loss optimization.This modification aim to achieve lightweighting while enhancing detection accuracy and speed.Compared to the original YOLOv5s model,the enhanced model exhibits a 14.8%increase in mAP,a reduction of 49.7%in parameters,a 40.7%decrease in model volume,and a 48.9%decrease in computational load.On the NVIDIA 3060 platform,the improved detection speed has surged by 48.8%,reaching 82 fps.Additionally,hardware acceleration using FPGA has been implemented for YOLOv5s.The optimized system achieves a detection frame rate of 45 fps while maintaining high precision and speed.This system is easily deployable and suits the demands of intelligent transportation systems,fulfilling the need for efficient real-time monitoring.