A real-time system for detecting indecent videos based on spatiotemporal patterns

In this paper, an efficient real-time system is proposed for detecting indecent video scenes, which usually contain periodically moving skin-colored objects. Spatiotemporal motion trajectories are efficiently extracted by the simple color-based region segmentation of a spatiotemporal pattern generated from an input video. By analyzing the vertical displacement of the spatiotemporal motion trajectory along the time axis of the spatiotemporal pattern, the trajectory is then converted to a one-dimensional (1D) signal. Feature vectors computed from the discrete Fourier transform (DFT) of the 1D signal and the colors near the spatiotemporal motion trajectory are used as the inputs for a classifier used to detect adult scenes. Experimental results show improvements in true positive and false alarm rates when compared to existing methods, and significantly reduced processing times.