The invention provides a vehicle low speed collision signal feature extraction method based on a waveform gradient. The adopted technical scheme is described as follows: a filtered waveform is subjected to gradient extraction; the minimum value Min and the maximum value of the gradient and the step size Step are set, and a histogram group number is calculated; counting is conducted according to histogram groups corresponding to a gradient scope to obtain the number of gradients within the gradient scope, namely, a gradient histogram; a mean value of all gradients is calculated; and the gradient histogram and the mean value of all gradients are taken as final gradient features. By calculating curve gradients, an original waveform is basically restored, and thus relatively clear learning features are provided for later machine learning.