基于改进YOLOv5s的老人跌倒识别算法研究

Research on the Elderly Fall Recognition Algorithm Based on Improved YOLOv5s

为应对人口老龄化加剧带来的老年人居家安全问题,提出了一种实时准确的跌倒识别算法。目前基于计算机视觉的跌倒检测主要存在跌倒类间相似性大、监控角度相对固定而跌倒角度多变、特征提取困难、难以联系图像依赖关系等问题。为此,采用同一动作8个拍摄角度的数据集,使模型能够充分学习到多种特征;将所有跌倒行为归为一类,总类别数为二分类;改进原网络的候选框、添加CC-Net、改进非极大值抑制算法,进行分类训练,并对训练好的模型进行对比评估。实验结果表明,该方法的平均准确率比YOLOv4和YOLOv5s模型分别提高了5.9%和2.6%,比SSD算法提高了11.1%,能够基本满足检测需求。

In order to deal with the problem of home safety caused by the aging population,a real-time and accurate fall recognition algorithm is proposed.At present,fall detection based on computer vision mainly has many problems,such as large similarity between fall classes,relatively fixed monitoring angle and variable fall angle,difficulty in feature extraction,and difficulty in contacting image dependencies.Therefore,the data set of 8 shooting angles for the same action is adopted,so that the model can fully learn multiple features.All fall behaviors are classified into one category,and the total number of categories is binary.The candidate frames of the original network are improved,CC-Net is added,and the improved NMS is for classification training.The trained models are compared and evaluated.Experiments show that the average accuracy of the proposed method is 5.9%and 2.6%higher than that of the original YOLOv4s and YOLOv5s models,respectively,and 11.1%higher than that of the traditional SSD series algorithms,basically meeting the detection requirements.