针对跌倒检测任务中复杂信息干扰和数据集缺乏导致模型精度不高的问题,设计一种高精度跌倒检测算法,降低模型参数的同时保持各种场景下的鲁棒性.该算法基于YOLOv5s改进,在骨干网络中使用Ghost module和解耦全连接注意力,以较低计算开销提升模型在光线变化、遮挡等干扰环境下的性能.在颈部层使用自适应感受野和空间通道混合注意力,提升神经元对不同尺度特征的适应性,应对人体形变、视角变化等干扰.引入EIoU损失函数,加速收敛提升训练精度.在公开数据集 Le2i Fall Detection Dataset 和 UR Fall Detection 上,精确率、召回率、mAP0.5 和mAP(0.5∶0.95)相比YOLOv5s分别提高4.0%,4.2%,2.9%和4.3%,参数量降低38.6%.该算法在多种应用场景下都保持较高检测精度,参数量较低,满足实际应用场景部署要求.
Improving YOLOv5s for Personnel Fall Detection Algorithm
Aiming at the problems of complex information interference and lack of data sets in the fall detection task,a high-precision fall detection algorithm is designed to reduce the number of param-eters while maintaining robustness in various scenarios.This algorithm is based on the improvement of YOLOv5s,using Ghost module and decoupling fully connected attention in the backbone network,re-ducing parameters while improving feature extraction capabilities,and improving the performance of the model in interference environments such as light changes and occlusions with low computing overhead.The adaptive receptive field and spatial channel hybrid attention are used in the neck layer to improve the adaptability of neurons to features of different scales and cope with interference such as human body de-formation and perspective changes.The EIoU loss function is introduced to accelerate convergence and improve training accuracy.On the public data sets Le2i Fall Detection Dataset and UR Fall Detection,the precision rate,recall rate,mAP0.5 and mAP(0.5∶0.95)are respectively improved by 4.0%,4.2%,2.9%and 4.3%compared to YOLOv5s,and the number of parameters is reduced by 38.6%.This algo-rithm maintains high detection accuracy in a variety of application scenarios and has a low number of pa-rameters,meeting the deployment requirements of actual application scenarios.
fall detectionYOLOv5sGhost moduleselective kernel convolutionEIoU