Curvelet transform sub-difference image for crowd estimation

Counting people and estimating their number is a fundamental task for many intelligent security systems, including CCTV systems and other visual surveillance research areas. This paper presents a new approach for crowd counting. The proposed method is independent of any background modelling or background subtraction techniques. Moreover, the new method is able to handle the perspective phenomena in a simple way. To do so, the estimation is determined by an enhanced version of a difference image. Every two sequential frames are used to extract a difference image. The curvelet transform is then applied to both frames. The information stored in every scale in the new sub-band images can be used as a source for different features after a customized inverse curvelet transform. Two different curvelet inverse transforms with three different features are used to evaluate the proposed counting algorithm; a Back Propagation Neural Network (BPNN) is used for crowd quantity predictions. The overall performance is measured over a UCSD benchmark dataset.