露天矿卡车司机驾驶行为Yolo-V5检测模型轻量级优化研究

Research on Lightweight Optimization of Yolo-V5 Detection Model for Driving Behaviors of Truck Drivers in Open-Pit Mine

针对Yolo-V5目标检测算法在露天矿应用中存在的模型拟合能力低、实时目标检测占用内存过大及设备配置要求高等缺陷,提出融合优化策略对Yolo-V5模型进行轻量级优化研究。在模型训练阶段,引入动量改进torch.optim.SGD()算法对模型训练权重进行优化,加快训练收敛速度,增强模型拟合度;模型训练后,在OpenVino环境下对Yolo-V5模型进行网络层剪枝和FP16量化,模型体积压缩75%,运算内存降低33.88%。通过实验室对比试验,得出如下结论:相较搭载GPU设备,采用融合优化策略的检测模型,在低配置CPU设备上推理FPS提升83.71%,平均推理时间降低45.11%,8类驾驶行为检测预警时间降低9.89%~82.21%,对吸烟、吃喝、手机检测精确度提升3.34%~10.00%。研究表明:融合优化策略实现了Yolo-V5轻量级优化的研究目标,实时目标检测摆脱了对GPU设备的依赖,为Yolo-V5目标检测进一步在露天矿应用推广打下了良好的研究基础。

Aiming at the defects of Yolo-V5target detection algorithm in open-pit mine application,such as low model fitting ability,excessive memory occupied by real-time target detection and high equipment configuration requirements,a fusion optimization strategy was proposed to carry out lightweight optimization research on Yolo-V5model.In the model training stage,the momentum improved torch.optimum.SGD()algorithm was introduced to optimize the training weight of model,accelerate the training convergence speed and enhance the fitting degree of model.After training,the Yolo-V5 model was pruned in the network layer and quantified by FP16in the OpenVino environment.The model volume was compressed by 75%and the computational memory was reduced by 33.88%.Through laboratory comparative experiments,the following conclusions are drawn.Compared with the detection model equipped with GPU equipment,using the fusion optimization strategy to infer FPS on low-configuration CPU equipment is increased by 83.71%,the average inference time is reduced by 45.11%,and the detection and early warning time of 8types of driving behavior is reduced by 9.89%-82.21%,and the detection accuracy of smoking,eating,drinking,and using mobile phone is improved by 3.34%-10.00%.The research shows that the fusion optimization strategy achieves the research goal of Yolo-V5lightweight optimization,and the real-time target detection gets rid of the dependence on GPU equipment,which lays a good research foundation for the further application and promotion of Yolo-V5target detection in open-pit mines.