基于改进YOLOv5s的跌倒行为检测

Fall Behavior Detection based on Improved YOLOv5s

摘要:

为了实现电厂人员跌倒行为的实时检测，防止跌倒昏迷而无法及时发现并救援的事件发生，针对跌倒行为检测实时性以及特征提取能力不足的问题，提出了一种改进YOLOv5s的跌倒行为检测算法网络：在YOLOv5s模型中引入SKAttention注意力模块，使得网络可以自动地利用对分类有效的感受野捕捉到的信息，这种新的深层结构允许CNN在卷积核心上执行动态选择机制，从而自适应地调整其感受野的大小；同时结合ASFF自适应空间融合，并在其中充分利用不同特征，又在算法中引入权重参数，以多层次功能为基础，实现了水下目标识别精度提升的目标；加入空间金字塔池化结构SPPFCSPC，大大缩短了推理时间。实验结果表明，相比于原始YOLOv5s，新网络在mAP平均精度均值方面提升了2.1%，查全率提升了16%。改进后的网络在感知细节和空间建模方面更加强大，能够更准确地捕捉到人员跌倒的异常行为，检测效果有了显著提升。

Abstract:

In order to achieve real-time detection of fall behavior among power plant personnel and prevent the occurrence of events that cannot be detected and rescued in a timely manner due to falls and coma, an improved YOLOv5s fall behavior detection algorithm network is proposed to address the issues of insufficient real-time detection and feature extraction capabilities. The introduction of SKAttention module in the YOLOv5s model enables the network to automatically utilize the information captured by effective receptive fields for classification.This new deep structure allows CNN to perform dynamic selection mechanisms on the convolutional core, thereby adaptively adjusting the size of its receptive field; By combining ASFF adaptive spatial fusion and fully utilizing different features, and introducing weight parameters into the algorithm, based on multi-level functions, the goal of improving the accuracy of underwater target recognition is achieved;The addition of spatial pyramid pooling structure SPPFCSPC greatly reduces inference time. The experimental results show that compared to the original YOLOv5s, the new network has improved the average accuracy of mAP by 2.1% and the recall rate by 16%. The improved network is more powerful in perception of details and spatial modeling, and can more accurately capture abnormal behaviors of people falling, significantly improving the detection effect.