基于头部运动轨迹和3D视觉的跌倒检测系统

为了实现自动识别人体跌倒行为的目的，采用深度摄像头同时获取彩色图像和深度图像，并对其进行校正；对深度图像采用背景差分法提取运动人体前景；利用彩色图像和深度图像的融合、肤色检测以及距离加权面积法进行头部的检测与定位；最后将头部运动速度作为判断跌倒的依据，对人体不同运动行为的头部速度进行采集，利用支持向量机对跌倒进行检测。以5个志愿者的步行、蹲下、坐下和跌倒等4类动作为实验样本，运用交叉比对确定最优化参数，最后采用本文提出的方法进行测试。实验结果表明，系统能够有效识别跌倒，总体识别率超过95％。其中，跌倒／步行行为分类识别率达100％；针对跌倒／蹲下和跌倒／坐下的系统平均识别率分别为94．4％和98．8％。

Fall detection system based on head moving trajectory and 3D vision

In order to achieve automatic recognition of human fall behaviour, the depth camera is used to obtain color images and depth images, which are corrected by the presented system. The foreground of moving human are extracted from the depth images by the background difference method, and then the head are detected with the fusion of color images and depth images, color detection and distance-weighted area method. Finally, by taking the head moving speed as the criterion of tumbling down, the head moving speed in human＇s various behaviors is acquired for detecting the fall by SVM. With the experimental samples, such as the walking, squating, sitting down and （ailing of five volunteers, some experiments were carried out and the optimal parameters were confirmed with cross-contrast. The results show that the system can effectively identify the human＇s tumbling down, the overall recognition rate is over 95%, and the classification recognition rate is： falling vs walking=100%, falling vs squating = 94, 4 %and falling vs siting down = 98.8%.