To solve the problem of high false negative rate in human fall detection, a detection algorithm based on D-S evidence theory is studied. Acceleration sensors and gyro sensors are built into a smartphone to obtain the three-dimensional motion data of the human body's arm movement. A three-step moving average filter is used to preprocess the three-dimensional raw data obtained from the two types of sensors. Then, the detection features consisting of movement range, inclination degree and rotation degree are extracted from the three-dimensional preprocessed data. The dynamic time warping method is used for local detection according to the three features. The local decision results act as evidences to be fused by the combinational rule of D-S evidence theory to get the final global detection result, in which each evidence is modified by the weight of evidence to avoid the evidence conflict problem. Experimental results show that the accuracy of the proposed algorithm is higher than that of other algorithms, which can effectively improve the detection performance.