A HYBRID MODEL FOR HUMAN ACTIVITY SURVEILLANCE USING DEEP LEARNING AND MOTION FLOW

Abstract

Human Activity Recognition from digital video is gaining popularity because of increasing crime rate worldwide. With the increase in crime rate, there is a massive increase in the installation of surveillance cameras. Monitoring these surveillance cameras on 24x7 basis, requires a lot of manpower. The accuracy of the system depends upon the activeness of the supervising entity. Sometimes, biased nature of the supervising person, fatigue issue results in the inaccurate decision. Therefore, there is a dire need of automated surveillance system, which can detect anomalous event correctly and in time. The problem of detecting anomalous event involves recognizing the human activity closely from the data collected through CCTVs. For the effective monitoring, an automated surveillance system is the need of time. In this work, a deep learning-based approach is used to classify the activities correctly from the video data. Collection of Image frames are considered for classification, as action in a single image frame may not be same in subsequent image frame. Therefore, in order to understand the human activities, analysis is made on sequence of image frames. In addition to spatial correlation present in the 2D images, temporal structure present in the video data is also considered.