Upper Body Detection in Unconstrained Still Images

The detection of upper bodies is fundamental to many computer vision tasks, with human pose estimation being our focus. Accurate upper body detection improves the robustness and reduces the search space for top-down as well as bottom-up approaches for pose estimation. This paper focuses on a particularly challenging task of detecting upper bodies from unconstrained still images. We propose a method that fuses the reliability of face detectors with the robustness of people detection based on HoG descriptors to improve the accuracy of upper body detection from monocular still images with cluttered background, poor illumination, motion blur and high-degree of occlusion. We compare the performance of the proposed method with six existing face and upper body detectors. Despite the relatively simple concept behind the proposed detector, it performed on par with the state of the art systems using challenging test images from the Buffy Stickmen v2.1 dataset.