Contextual Multi-scale Region Convolutional 3D Network for Anomalous Activity Detection in Videos

Abstract

The importance of the surveillance camera has been greatly increased as a result of increase in the crime rates. The monitoring of the videos recorded is once again challenging. Thus, there is an incredible interest for intelligent surveillance system for security purposes. The proposed work aims at a anomalous activity detection in video using the contextual multi-scale region convolutional 3D network (CMSRC3D) for. The work is deal with different time scale ranges of activity instances; the temporal feature pyramid is used to represent normal and abnormal activities of different temporal scales. In the temporal pyramid are 3 levels, that is short, medium and long for video processing. For each level of the temporal feature pyramid, an activity proposal detector and an activity classifier are learned to detect the normal and abnormal activities of specific temporal scales. Most importantly, the entire system model at all levels can be trained endwise. The activities detected are classified as the normal and the abnormal using the activity class. In the project CMSRC3D, each scale model uses convolutional feature maps of specific temporal resolution to represent activities within specific temporal scale ranges. More importantly, the whole detector can identify anomalities with in all sequential ranges in a single shot, which makes it very efficient for computation. And also provide if any anomalous activities detected in video, we give alert to avoid future mishappening.