A real-time cascade pedestrian detection based on heterogeneous features

In this paper, we are introducing a real-time pedestrian detector in traffic scenes. To improve the running efficiency of the whole detection system, a cascade framework is implemented. Three heterogeneous types of features are used, including aggregate channel feature (ACF), the responses of self-similarity and checkerboard filters on ACF. Based on the observations that 1) these three types of features take different complexities to compute, and 2) most of the false positives are rejected in the former cascade stages, we propose to divide the cascade into 3 parts, and each type of features is selected in each cascade part according to the feature complexity. The ACF features are computed for the whole image before any cascade stage, and the other two types of features are computed when they are needed at specific stages. This strategy significantly outperforms ACF-only features with only a little loss of speed. Finally, we achieve a pedestrian detector with running speed of 10 frames per second on a 640×480 image.