One of the health issues in elderly people is the injury from the fall. Some of these injuries might lead to deaths. Thus, a good fall detection algorithm is needed to help reducing a rescuing time for a helper. In this paper, we develop a fall detection algorithm using the linear prediction model with a tri-axis accelerometer. We test the algorithm with the data set that have 11 activities (standing, walking, jumping, falling, running, lying, sitting, getting up (from lying to standing or from sitting to standing), going down (from standing to sitting), accelerating and decelerating) from 17 subjects. The result shows that we can detect all fall activities in both training and blind test data sets with precisions of 90.72% and 93.69%, respectively. The result also shows that we can detect 89.77% and 93.27% of other activities correctly. Although, there are some false alarms, the false alarm rate is small. © 2013 IEEE.