In recent years, the aging of population is one of the problems that many countries need to face. Along with the increasing proportion of elderly people living alone, there are more indoor but fatal accidents. Fall is one of these common and dangerous accidents for the elderly. Thus timely rescue after falls becomes particularly important, especially for elderly people who live alone. With the development of computer vision technology and the popularity of home surveillance, the fall detection algorithm based on video analysis provides a good solution to this problem. In this paper, we propose a new fall events detection algorithm. Our algorithm gets sub-motion history image by mapping faster R-CNN detected bounding boxes to motion history image, then extracts histogram of oriented gradient features, and finally uses support vector machine for fall classification. Proved by experiment, Our approach achieves very high recall rates and precision rates in a dataset of realistic image sequences of simulated falls and daily activities.