改进型YOLOv8融合关键点的跌倒检测算法

Fall detection algorithm based on the improved YOLOv8 combined with key points

针对传统跌倒检测算法在特征提取不充分、跌倒判决方法单一以及实时性不强的问题,提出一种改进型YOLOv8结合人体骨骼关键点的跌倒检测算法。首先,算法通过ShuffleNetV2网络替换原有YOLOv8的Darknet-53主干网络,在颈部增加混合注意力机制(Shuffle Attention,SA),使得网络能够更好地提取人体的行为特征,实现人体静态姿势匹配。其次,分析人体动态行为的骨骼关键点位置变化信息,将人体质心下降速度、人体躯干与地面间的夹角变化速度和人体的高宽比三者共同作为跌倒行为的判决依据,提高跌倒判决的准确率。实验结果表明,该算法在COCO Key Points数据集上的检测精度、F1值和mAP50值分别为78.3%、67.9%和70.0%,在UR Fall Detection、Fall Detection Datasets和自建数据集上的检测准确率分别为95.85%、92.8%和96.52%,在区分日常生活行为和跌倒行为方面优于传统算法。

To solve the problems of insufficient feature extraction,a single fall detection method,and weak real-time performance of traditional fall detection algorithms,an improved YOLOv8 fall detection algorithm combined with human skeleton key points is proposed.First,the backbone network of YOLOv8 is replaced by a ShuffleNetV2 network,and the mixed attention mechanism(Shuffle Attention,SA)is added in the neck,so that the model can extract the behavioral characteristics better and realize the static posture matching of a human body.Second,by analyzing the information on position change of skeletal key points,the decline speed of the center of mass,the angle speed between the trunk and the ground and height-to-width ratio of the body are taken as the basis of the fall behavior to improve the accuracy of fall judgment.Experimental results show that the algorithmic accuracy,F1 value,and mAP50 value on COCO Key Points datasets are 78.3%,67.9%,and 70.0%respectively,that the algorithmic accuracy is 95.85%,92.8%and 96.52%on UR Fall Detection,Fall Detection Datasets and self-built datasets,and that the proposed algorithm outperforms the traditional algorithm in distinguishing daily life behavior and falling behavior.