Objective To develop a kind of algorithm for fall detection based on human acceleration. Methods From September to November, 2017, six healthy postgraduates participating in the experiment completed 13 acts of falls and eleven of activities of daily life. The information of activities was collected through two acceleration sensors, 81 acceleration features were extracted from each sensor, and were reduced dimension through principal component analysis. K-nearest neighbor was used to detect the falls and activities of daily living. Results The sensitivity of fall detection was 100%, the specificity was 99.76%, and the detection time was 216 ms. Conclusion The algorithm of multi-feature fusion of human body acceleration and K-nearest neighbor is accurate and timely.