Object detection using deformable part model in RGB-D data

Abstract

Object detection is an important research field in computer vision, previous studies have focused on detection in standard RGB images, with the popularity of motion-sensing camera, such as Kinect, combining RGB and depth data in object detection becomes a new challenge. In this paper, the object model is analyzed thoroughly with the trait of depth information, a criteria CIR (Contour Intensity Ratio) is given out to measure the contour strength of model, and then the advantage of depth data in encoding the contour strength is illustrated. The Deformable Part Model (DPM) [1] and Histogram of Oriented Depths (HOD) [2] are combined to a new RGB-D joint detection strategy, and the 4 rules of depth filtration after detection as well as part-class depth filtration are put forward, the important role of depth information for object detection is confirmed by human detection experiment.