Pedestrian Video Data Abstraction and Classification for Surveillance System

Abstract

In this study, we have developed abstracted pedestrian behavior representation and classification method for pedestrian video surveillance system. An effective intelligent surveillance system can be constructed if the high-resolution surveillance image information is efficiently summarized. The motion of the pedestrian is represented by a multi-layer grid map using a detector and a tracker. A normal pattern and anomalous pattern database were constructed and classified using the CNN classifier. With the abstracted pedestrian data and CNN network, the abnormal situation can be detected up to recall 92.0%, precision 99.9%.