Improving counterflow detection in dense crowds with scene features

This paper addresses the problem of detecting counterflow motion in videos of highly dense crowds. We focus on improving the detection performance by identifying scene features - that is, features on motionless background surfaces. We propose a three-way classifier to differentiate counterflow from normal flow, simultaneously identifying scene features based on statistics of low-level feature point tracks. By monitoring scene features, we can reduce the likelihood that moving features' point tracks mix with scene feature point tracks, as well as detect and discard frames with periodic jitter. We also construct a Scene Feature Heat Map, which reflects the space-varying probability that object trajectories might mix with scene features. When an object trajectory nears a high-probability region of this map, we switch to a more time-consuming and robust joint Lucas-Kanade tracking algorithm to improve performance. We evaluate the algorithms with extensive experiments on several datasets, including almost three weeks of data from an airport surveillance camera network. The experiments demonstrate the feasibility of the proposed algorithms and their significant improvements for counterflow detection.