基于光流重构和深度姿态特征的行人跌倒检测方法

随着我国老龄化人群数量的增长，老年人实时行为轨迹的跟踪检测成为当前智慧社区研究的热点。本文针对现有行人跌倒检测算法在面临电动车流和行人影子的干扰而导致识别准确率不高的问题，提出了一种基于优化yolov5的路边行人跌倒检测方法。首先基于视频抽帧的方式将原始数据输入到yolov5网络进行监控行人视频数据的预处理，实现行人背景重构；然后提取光流和anchor框作为其运动特征；最后对此特征进行判定，进行信息融合的跌倒检测网络进行跌倒特征识别，并在不同帧序列和不同背景下进行对比试验。本文所提出的方法在行人跌倒数据集Multiple cameras fall和Le2i上进行了实验，结果表明本文的算法在基于电动车流和行人影子干扰场景下较传统方法在准确率和召回率上分别提升了9%和10%。

Pedestrian fall detection method based on optical flow and posture network

With the growth of the aging population in China, the tracking and detection of the real-time behavior trajectory of the elderly have become a hot spot in the current smart community research. In this paper, a yolov5 roadside pedestrian fall detection method based on optical flow and posture is proposed to solve the interference of electric vehicle flow and pedestrian shadow. Firstly, the original data is input to yolov5 network by video frame extraction to preprocess the video data of pedestrian monitoring and realize the reconstruction of pedestrian background; Then the reference frames of optical flow and pedestrian pose are extracted as their motion features; Finally, these features are judged and the fall detection network based on information fusion is used to recognize the fall feature. The comparative experiments are carried out in different frame sequences or backgrounds. The proposed method is conducted on the pedestrian fall public datasets (such as multiple cameras fall and le2i). The results show that the accuracy and recall are improved by 9% and 10% respectively compared with the traditional methods.