Crowd Panic Detection Using Autoencoder with Non-uniform Feature Extraction
Michael D. George, C. V. Bijitha, Babita Roslind Jose · 2018
This paper presents a crowd panic detection method based on an autoencoder that uses motion features extracted from non-uniform spatio-temporal region. The autoencoder is fed with recent motion based feature called Histogram of Optical Flow Orientation and Magnitude (HOFM). This feature is extracted from a non-uniform spatio-temporal region that considers the surveillance camera position with respect to the scene. The autoencoder is trained with such features from scenes that involve normal crowd behaviour. The detection of a crowd panic situation is based on the fact that the trained autoencoder will struggle to reconstruct the features extracted from a crowd panic scene. The proposed crowd panic detection methodology is tested on three crowd panic sequences that are part of the LV dataset.