基于人体骨骼点改进的局部椭圆拟合跌倒检测方法
Improved Fall Detection Method Based on Local Elliptical Fitting of Human Skeletal Points

老年人跌倒所造成的生理和心理问题是对健康的重大危害，也对独立生活造成极大障碍。为了解决传统的使用几何形状和前景检测对人体外轮廓特征进行跌倒检测的不稳定性、难以区别相似活动以及算法过于复杂等问题，本文提出了一个基于人体骨骼点改进的局部椭圆拟合跌倒检测方法。首先，使用OpenPose算法提取人体骨骼点位置坐标并筛选出用于拟合的骨骼点。然后通过骨骼点坐标计算拟合出椭圆轮廓，提取椭圆的长短轴之比、方向角和中心点垂直方向速度，融合成一个运动特征。最后，经过一个CNN对运动特征加以训练，用于跌倒行为检测判断。试验结果表明，本文方法相较于现有方法，有效克服了人体外几何轮廓的不稳定性，从而大大提高检测的准确率。
The physiological and psychological problems caused by falls in the elderly are a significant threat to health and pose a great barrier to independent living. To address the instability of traditional fall detection methods that use geometric shapes and foreground detection for human contour features, which struggle to distinguish similar activities and are overly complex, an improved fall detection method based on local elliptical fitting of human skeletal points is proposed. Initially, the OpenPose algorithm is used to extract the coordinates of human skeletal points and select those suitable for fitting. Then, an ellipse is fitted based on the skeletal points in each frame, extracting the ratio of the major to minor axis, the orientation angle, and the vertical velocity of the center point to create a motion feature. Finally, a Convolutional Neural Network (CNN) is used to train these motion features for the classification and judgement of fall behavior. Experimental results show that this method overcomes the instability of human body geometric contours compared to existing methods and improves the detection rate.