基于改进YOLOv5与边缘计算的智能猪只盘点

Research on intelligent pig inventory based on improved YOLOv5 and edge computing

针对生猪检测盘点算法在复杂场景中泛化能力弱,猪只遮挡误差较大,且推理速度慢,不易在边缘设备部署的问题,提出一种改进的YOLOv5算法,并将其部署在NVIDIA Jetson Xavier NX边缘设备上进行猪只盘点。采用半监督伪标签方式的数据增强策略,提高泛化能力;在YOLOv5主干网络中加入ECA注意力机制实现跨通道的信息交互,提高模型的精度及推理速度;采用SPPCSPC模块替换SPP模块,在原有基础上引入残差,更好地引导梯度流的传递,提高模型的精度;采用SIoU损失函数,减少回归框的“震荡”,降低盘点误差,提高模型推理速度。在实际场景猪只数据集上的实验结果表明,与改进前相比,平均绝对误差和均方根误差分别降低0.69、0.65,平均精度均值和召回率分别提高2.02%、2.75%,帧率提高6.53 fps。该算法可以有效减小猪只遮挡产生的误差并提高算法推理速度,实现了基于边缘设备的实时猪只盘点。

To solve the issues of weak generalization ability,large pig occlusion error,slow reasoning speed,and the difficulty of deployment on edge devices in complex scenes,an improved YOLOv5 algorithm is proposed and deployed on the NVIDIA Jetson Xavier NX edge device for pig inventory.The algorithm adopted data augmentation strategy of the pseudo-label semi-supervised learning to improve the generalization ability.ECA attention mechanism was added to the YOLOv5 backbone network to enable cross-channel information interaction,which improved the accuracy and inference speed of the model.The SPPCSPC module was used to replace the SPP module,and residual error was introduced to better guide the transfer of gradient flow and enhance the accuracy of the model.The SIoU loss function was used to reduce the"oscillation"of the regression frame,decrease the inventory error,and improve the reasoning speed of the model.The experimental results on the actual scene pig dataset demonstrated that the mean absolute error and root mean-square error decreased by 0.69 and 0.65,respectively,after the improvements.Additionally,the mean average precision and recall increased by 2.02%and 2.75%,respectively.Moreover,frame per second increased by 6.53.The algorithm effectively reduced the error caused by pig occlusion and improved the reasoning speed.Furthermore,it enabled real-time pig inventory based on edge devices.