Effective strategies for automatic detection of abnormal activities in surveillance video streams

Abstract

In Computer Vision, the surveillance is monitoring of activities, behaviour and other changing information. Every year the automated visual analysis of behaviour provides some key building blocks towards an intelligent vision system. The capacity to perceive people and their actions by vision is a key for a machine to interface robustly and easily with a human computer interacting world. One of the most applications of visual analysis is anomaly detection in human activities. In video processing, anomaly is generally considered as a rare occurring event. An anomaly can appear in various forms they represent the various levels of human safety issues. The detection and tracking of abnormal activities in surveillance have inspired an increasing level of concentration in computer vision. Therefore, this research work aims to propose a new category of abnormality detection techniques. The main aim is to design the model that detects the abnormality in the human behavioureither static dangerous objects (e.g., luggage or bomb abandoned in public places) and dynamic abnormal behaviours (e.g. fighting) and to recognize the crowd event in the video sequence based on features extracted and the movement of motion vectors in video streams. This results in effective identification in abnormal behaviours in a complex public environment that ensures the security and safety. The approaches used for effective automatic detection of abnormal activities in video streams are as follows: 1. Hash Based Abandoned Object Detection (HAOD) 2. Cue Based Human Behaviour Abnormality Detection System (CHBDS) iv 3. A Deep Learning Framework for Abnormal Event Detection (DLAD) In this work the first approach is Hash based abandoned object detection, which is proposed to detect the abandoned objects in the crowded areas. The main intention of this research work is to propose an intelligent model for identifying the abandoned objects in crowded areas using improved hashing techniques with the combination of short-term and long-term blob logic, a