Anomaly detection in real time CCTV streams

Abstract

Anomaly detection in video data has been a challenge always. After the introduction of many state-of-art designs, this still poses a challenge as those systems may fail to work in all types of environments. Even though many supervised methods claimed to have some good results in this domain, supervised systems may not be suitable for all the contexts such as in an open area, any type of anomaly can occur and it can be very di cult to train a system in a supervised manner to identify an unanticipated anomaly. On the other hand, it would be di cult for the user to annotate data each time when they change the context under surveillance for the device. Thus the ultimate solution should be an unsupervised solution with a appreciable accuracy. Recently deep learning techniques have emerged in many areas of computer science based solutions and so it is involved for anomaly detection tasks also. In this research, deep learning techniques are involved to solve the problem of video stream based anomaly detection of crowds.