Violence Detection from Videos using HOG Features

Advancement of modern technologies over the last few decades drive action recognition to become an essential topic in surveillance systems. Violent action poses a significant threat to our freedom and social security. Therefore, it is crucial to classify violence in surveillance scenarios like in a prison, gymnasium, car parking, and so on. In this work, a novel method is proposed to identify violence in different circumstances. Initially, some frames are selected from each video clip using image subtraction and averaging techniques and Histogram of Oriented Gradient (HOG) is applied to extract lower level features. Finally, Support Vector Machine (SVM), Logistic Regression, Random Forest, Linear Discriminant Analysis (LDA), Naïve Bayes and K-Nearest Neighbors (KNN) are used for the classification purpose. The system is able to achieve the highest accuracy rate of 86% while using Random Forest classifier. The experimental result on the benchmark dataset shows significant accuracy and an improvement over the previously proposed methods.