Detecting Violent Arm Movements Using CNN-LSTM

With the increasing rate of surveillance camera utilization in household residences, intelligent security systems capable of automatically identifying violent activity has already become a significant research issue in recent times. Indoor violence mainly comes in the form of punching, beating and scuffling which primarily involve arm movements. In this paper, an end-to-end deep learning framework for detecting violent arm movements has been proposed. To consider the motion information of the violent movements from the surveillance video clips, a combination of Convolutional Neural Network (CNN) and Long-Short-Term-Memory (LSTM) model is used. The CNN model extracts the spatial features which are passed to the LSTM model for temporal feature learning. The spatio-temporal features are then fed to the Softmax classifier for binary classification and outputs violent or non-violent label. Experimental results obtained after evaluating the proposed model on the benchmark Hockey Fight dataset exhibit higher classification accuracy than some of the existing state-of-the-art methods, thus, proving superior performance in the context of detecting violent arm movements.