Abnormal Action detection in video surveillance

Abstract

The growing number of anomalies happening in indoor and outdoor environments calls for accurate and robust action recognition systems. These anomalies could vary from theft, destruction of public property or even fighting innocents. The aim of this paper is to introduce a new algorithm based on machine learning paradigm to detect human actions and to label them as normal or abnormal. The algorithm starts by testing two different human detectors, cascade object detector and Faster Region Convolutional Neural Network for Human Detection (FRCNNHD). Both detectors were trained using widely available datasets. Afterwards, detected human figures are ex-tracted to form a video patch that represents human motion. For action recognition, we applied the Motion History Image to extract static features of motion. The actions are then classified using the Support Vector Machine (SVM). Finally, actions with low recognition confidence are labeled as “abnormal actions”. Experimental results on two datasets show the accuracy of our algorithm on learned actions.