Eliminating false positive results to effectively analyze anomaly changes in violent videos

Abstract

Due to the increasing incidents of crime and violence it is important to develop technology to automatically detect the presence of violence in security camera images. Although law enforcement agencies have sufficient images they do not have the human resources to analyze them and detect violence in a timely manner. In these video footage there are examples of false recognitions that are labeled as normal in some frames while the abnormal event continues. In this study we aim to identify the start and end frames of the event with minimum error in indoor or outdoor violent camera images. For this purpose firstly a model is created to enrich the sequential video frames containing violence by using MixUp data augmentation method for limited training datasets so that the system can learn more features and thus increase the training performance. Secondly with another proposed method more effective video analysis is realized by filtering the frames containing false positives in the outputs obtained from a deep learning-based system. Experimental results show that the proposed method achieves a remarkable success rate reaching 976% F-1 score and 954% IoU score values. In this way false positives are significantly reduced and the start end and action times of violent events that continue for more than one second in consecutive frames can be accurately detected.