Crowd counting using accumulated HOG

People count is an important indicator in video surveillance. Due to the overlapping objects and cluttered background, counting people accurately in actual crowded scene remains a non-trivial problem. Existing regression-based methods either learn a single model mapping the global feature to people count, or estimate localized count by training a large number of regressors. In this paper, we present an intermediate approach using the accumulated HOG feature. Our approach is able to capture the spatial difference of crowd structure and does not need to train a large number of regressors. Contrast to the low-level features existing regression-based methods generally use, the accumulated HOG feature is more robust. Extensive evaluations have been done on five benchmark datasets in the field of crowd counting, which demonstrate the robustness and effectiveness of our approach. In particular, the processing speed is fast enough to be applied to practical applications.