Hand detection using multi-resolution HOG features

This paper presents a novel feature based multi-resolution framework for hand detection. In the algorithm, Histogram of Oriented Gradient (HOG) is used for basic feature representation. To avoid the time consuming down-sampling procedure, features of increasing resolutions are directly extracted from the proposed Gradient-Orientation Image (GOI) under decreasing scales. For efficient detection, cascade is trained by letting the high stage use more discriminative high resolution features. The earlier stages can reject a large quantity of negatives by just using computational cheap low resolution features, and the later stages will carefully diagnose a small number of remaining candidate regions using powerful and expensive high resolution features. Extensive experiments are implemented to demonstrate its improvement and efficiency under complicate scenarios.