Early Recognition of Suspicious Activity for Crime Prevention

Abstract

Automatic identification and early prediction of suspicious human activities are of significant importance in video surveillance research. By recognizing and predicting a criminal activity at an early stage, regrettable incidents can be avoided. Initially, an action recognition framework is developed for identifying the suspicious actions using interest point based 2D and 3D features and transform based approaches. This is subsequently followed by a novel approach for predicting the suspicious actions for crime prevention in real-world scenario. The prediction problem is formulated probabilistically and a novel approach that employs the mixture models for prediction is introduced. The developed system yields promising results for predicting the actions in real-time.