一种融合运动特征的行人跌倒检测算法
2025,33(2):212-220
摘要：针对传统跌倒检测算法局限于单帧图像判断而未充分利用行人时序运动特征的问题，提出了一种融合运动特征的行人跌倒检测算法；在YOLOv8n的主干网络中引入ECA注意力机制，在Neck中用GSConv代替常规卷积提高行人检测的精度；为了避免多行人之间的特征发生混淆，使用DeepSort目标跟踪算法进行行人跟踪，独立提取行人的运动特征；分类阶段设置一个固定长度的队列来存储行人运动特征值，利用提前训练好的1DCNN进行跌倒识别。实验证明，在VOC2007测试集上，改进后的YOLOv8n模型比改进前的mAP增加1.5%，参数量减少2.97%；跌倒检测算法在UR Fall数据集上实现了97.8%的准确率；在自建的多行人数据集上实现92.91%准确率。
关键词：深度学习;YOLO;目标检测;运动特征;跌倒检测
A Pedestrian Fall Detection Algorithm Integrating Motion Features
Abstract：Aiming at the problem that traditional fall detection algorithms are limited to single-frame image judgment and do not fully utilize the temporal motion characteristics of pedestrians, a pedestrian fall detection algorithm that integrates motion features is proposed; the ECA attention mechanism is introduced into the backbone network of YOLOv8n, and GSConv is used instead of conventional convolution in Neck to improve the accuracy of pedestrian detection; in order to avoid collisions between multiple pedestrians When features are confused, use the DeepSort target tracking algorithm for pedestrian tracking and independently extract pedestrian motion features; in the classification stage, a fixed-length queue is set up to store pedestrian motion feature values, and 1DCNN trained in advance is used for fall recognition. Experiments have shown that on the VOC2007 test set, the improved YOLOv8n model increased mAP by 1.5% and reduced the number of parameters by 2.97% compared with the pre-improved model; the fall detection algorithm achieved an accuracy of 97.8% on the UR Fall data set; in the self-built Achieved 92.91% accuracy on multi-pedestrian data set.
Key words：deep learning; YOLO; target detection; motion features; fall detection
收稿日期：2023-12-12
基金项目：山东省重点研发计划(No.2019GSF111054),国家自然科学基金(62073196,U1806204)
