ANOMALOUS HUMAN ACTIVITY RECOGNITION FROM VIDEO SEQUENCES USING BRISK FEATURES AND CONVOLUTIONAL NEURAL NETWORKS

Abstract

In the present world, smart video surveillance system is essential for people to utilize distinctive sorts of security system to keep their property safe from unapproved person’s entry. A security system helps individuals to feel safe while they have to travel or go out of their home for work purposes and others. The current security systems against robbery are costly as a lot of money must be paid to the administration supplier to store the recorded video despite the very fact that there's no human movement recognized. The solution to this problem is an intelligent video surveillance system that can detect anomalous human activities automatically. This eventually minimizes the specified space for storing and makes the system cost-effective. A typical smart video surveillance system consists of two steps namely, feature extraction and classification.  In the proposed research work the Binary Robust Invariant Scalable Points (BRISK) features are extracted from the given videos during the feature extraction step. The obtained features are given as input to 3D Convolutional Neural Network (CNN) which classifies the videos as anomalous and normal based on the features. The present work uses University of California (UCF) dataset which contains thirteen video sets with one normal and twelve abnormal activity videos to evaluate the proposed method. The proposed BRISK and CNN based method provides a 100% training and 87.5% testing accuracy which is high compared to the existing methods that are based on shallow neural networks in detecting anomalous activities.