Salient motion detection in crowded scenes

To reduce cognitive overload in CCTV monitoring, it is critical to have an automated way to focus the attention of operators on interesting events taking place in crowded public scenes. We present a global motion saliency detection method based on spectral analysis, which aims to discover and localise interesting regions, of which the flows are salient in relation to the dominant crowd flows. The method is fast and does not rely on prior knowledge specific to a scene and any training videos. We demonstrate its potential on public scene videos, with applications in salient action detection, counter flow detection, and unstable crowd flow detection.