Abstract:
There is a need for automation in security systems and forensic investigations. Normally, for surveillance or for forensic investigation a person has to constantly monitor or watch the screen to look for any anomaly that might occur in the vicinity in which that camera is deployed or from where that video came. In reality, it is cumbersome and inefficient as it leads to human error. Due to minimal research in this field there are only a few working commercial software that would allow us to detect anomaly automatically. We aim to create an application that would allow us to automatically detect an anomaly in a required video eliminating the tedious labour one has to put in when viewing video samples. This provides a more vigilant surveillance aiding in a safer environment around a vicinity
We have used a comprehensive and innovative strategy to detect and localize anomalous objects, actions and activities from a scene. Motion of the moving object, size and the texture present in its neighbouring pixels are used as key features needed to make the decision whether the object is an anomaly or not. Analysing each feature helps us dealing with crowded scenes where many traditional algorithms fail to work.
The algorithm shall ignore background dynamics by focusing on the object of interest which are in motion. No hefty amount of training is required in order to yield good results. Comparison with state of the art algorithms have been made which gives us promising results in accuracy and speed. The algorithm could be enhanced and be used for various other applications like background estimation or robust foreground estimation. Anomalous event can be defined as any event that is different from what has been observed beforehand. The basis of this definition, what is an anomaly helps us detect anomalies in different scenarios and context. Detailed information of the working of the algorithm follows throughout the paper.
