Statistical t+2D subband modelling for crowd counting

Counting people automatically in a crowded scenario is important to assess safety and to determine behaviour in surveillance operations. In this paper we propose a new algorithm using the statistics of the spatio-temporal wavelet subbands. A t+2D lifting based wavelet transform is exploited to generate a motion saliency map which is then used to extract novel parametric statical texture features. We compare our approach to existing crowd counting approaches and show improvement on standard benchmark sequences, demonstrating the robustness of the extracted features.