Fight Detection in Crowd Scenes Based on Deep Spatiotemporal Features
Alaa Atallah Almazroey, Salma Kammoun Jarraya · 2020
Over the last few decades, remarkable growths of distributing surveillance cameras have been noticed in privet and public facilities. Therefore, with a growing demand for security and to ensure public safety. An intelligent automated approach is essential for on-spot violence and fight detection as it would save the time and cost of manual fight detection from monitor screens. In this research, we propose a supervised deep learning-based approach to detect fight actions from crowd video scenes. By extracting the keyframes from video frames using the cosine similarity algorithm, using them to compute optical flow values of the magnitudes, orientation and velocity. Afterward, using these values to construct four 2D templates, which are supplied to the pre-trained network to extract deep spatiotemporal features. Finally, applying the Neighborhood Component Analysis (NCA) feature selection method, and the Vector Support Machine (SVM) classifier, to generate a model that able to identify fighting behavior from crowd video scenes. Three different public datasets, the Hockey dataset, Movies dataset and Violent-Flows dataset, were used to evaluate the proposed method. The efficiency of the proposed method outperformed other state-of-the-art methods in terms of accuracy and the required time for fight detection to be executed.