Using SVM for Efficient Detection of Human Motion

Abstract

This paper presents a method for detection of humans in video. Detection is here formulated as the problem of classifying the image patterns in a range of windows of different size in a video frame as "human" or "non-human". Computational efficiency is of core importance, which leads us to utilize fast methods for image preprocessing and classification. Linear spatio-temporal difference filters are used to represent motion information in the image. Patterns of spatio-temporal pixel difference is classified using SVM, a classification method proven efficient for problems with high dimensionality and highly non-linear feature spaces. Furthermore, a cascade architecture is employed, to make use of the fact that most windows are easy to classify, while a few are difficult. The detection method shows promising results when tested on images from street scenes with humans of varying sizes and clothing.