Fast human detection from videos using covariance features

In this paper, we present a fast method to detect humans from videos captured in surveillance applications. It is based on a cascade of LogitBoost classifiers relying on features mapped from the Riemanian manifold of region covariance matrices computed from input image features. The method was extended in several ways. First, as the mapping process is slow for high dimensional feature space, we propose to select weak classifiers based on subsets of the complete image feature space. In addition, we propose to combine these sub-matrix covariance features with the means of the image features computed within the same subwindow, which are readily available from the covariance extraction process. Finally, in the context of video acquired with stationary cameras, we propose to fuse image features from the spatial and temporal domains in order to jointly learn the correlation between appearance and foreground information based on background subtraction. Our method evaluated on a large set of videos coming from several databases (CAVIAR, PETS, ...), and can process from 5 to 20 frames/sec (for a 384x288 video) while achieving similar or better performance than existing methods.