Abstract: In order to solve the problems of the instability of the traditional method of using geometric features to detect falls, and  the difficulty to distinguish some similar activities, a fall detection method based on convolution neural network (CNN) and the motion features of the elliptical contour of  human body is proposed. First, this method uses the Gauss mixture model to detect the human target and find out the minimum external elliptical contour. Then, the three features of the long and short axis ratio, the orientation angle and the vertical velocity of the human body’s centroid extracted in each frame’s elliptical contour, are fused into a motion feature based on time series. Last, a shallow CNN is then trained to detect falls and distinguish some similar activities. Experiment results show that our method overcomes the instability of geometric features and therefore enhances the detection rate compared with the existing methods.

Key words:  fall detection, convolution neural network, human elliptical contour, time series, motion features