Violent Scene Detection based on Multiple Instance Learning and 3D Histogram of Oriented Gradients.

In order to take advantage of the logical structure of video sequences and improve the recognition accuracy of the violence videos, a novel violent scene detection method based on 3D Histogram of Oriented Gradients (HOG3D) and Multi-Instance Learning (MIL) is proposed. Firstly, HOG3D was extracted from video clips. Then, K-means clustering algorithm was implemented to generate visual words vocabulary, and Bag of Visual Words (BoVW) model was used to construct the final feature for video frame sequence. Next, the violent scene detection issue was formulated as a MIL problem, and an instance clean method was proposed to remove noise in sample data. With experimental evaluations on the well-known Hockey and Movies benchmarks, the results demonstrate that Citation-kNN is more suitable than the other two tested MIL methods named Axis-Parallel hyper Rectangle (APR) and mi-SVM for violent scene detection. And the proposed scheme obtains very competitive results: 92.7% on Hockey and 98.8% on Movies respectively in terms of mean average precision. And it outperforms the stateof-the-art on Hockey dataset and matches the best known accuracy on Movies, achieving the best balanced accuracy compared with all other methods.