Panic Behavior Detection using Unsupervised Machine Learning Techniques: A comparative study

Crowd-scenes analysis is a hot research topic in computer vision nowadays. Over past years, many cases of crowd disasters happened overall the world. Real-time and fast panic behavior detection is crucial to quickly alert the crowd managers so they take appropriate precautions. Machine learning, mainly, unsupervised learning, has made great progress in abnormalities detection. In another hand, the Optical Flow (OF) feature forms the core of numerous detection models as it has proven its capability to represent and interpret complex motion for a long time. In this paper, we review the recent existing works in the field. The empirical comparison of common unsupervised classifiers using OF is provided to investigate their performance for panic detection in terms of correct detection, false detection, and frame rate. The investigation on these performance metrics reveals that the Autoencoder has the best detection performance in most of the dataset scenarios, while K-mean and hierarchical clustering are the fastest.