A waveform feature recognition algorithm based on SVM is introduced, and how the algorithm is applied to human body acceleration waveform recognition is described. First, a waveform decision model is established by using LIBSVM, and a training set is established to train and cross-verifying the accuracy of the model by using a waveform of falling and normal motion. By adding the sliding observation window on the continuous waveform, the decision model can make judgment based on the waveform segment in the window, in this way the falling waveform is detected in real time, and can be divided from the waveform of normal motion such as running and walking. In that case of misjudgment, the training set can corrected in time, the fall determination model is continuously trained, and the judgment accuracy is continuously improved.