Crowd Pedestrian Counting Considering Network Flow Constraints in Videos.

A quadratic programming method with network flow constraints is proposed to improve crowd pedestrian counting in video surveillance. Most of the existing approaches estimate the number of pedestrians within one frame, which result in inconsistent predictions in temporal domain. In this paper, firstly, we segment the foreground of each frame into different groups, each of which contains several pedestrians. Then we train a regression-based map from low level features of each group to its person number. Secondly, we construct a directed graph to simulate people flow, whose vertices represent groups of each frame and edges represent people moving from one group to another. Then, the people flow can be viewed as an integer flow in the constructed directed graph. Finally, by solving a quadratic programming problem with network flow constraints in the directed graph, we obtain a consistent pedestrian counting. The experimental results show that our method can improve the crowd counting accuracy significantly.