An RGB-D Descriptor for Object Classification

One of the main and active research areas in computer vision is the object detection which has various applications including image retrieval, video surveillance, robotics, etc. The main problem of object detection is, detecting instances of semantic objects of predefined classes (such as pedestrians, faces, or cars) in 2D images and videos. As 2D images of the objects include information about object appearance, most of the methods rely on pattern detection algorithms using appearance-based or feature-based techniques. Although the availability of 3D image data by using inexpensive depth cameras has made the problem more tractable, many researchers still tend to use similar concepts applied to the 2D instance problem. In this paper, we aim to develop a 3D descriptor that exploits the information in 3D data to address the many difficulties associated with object detection. This method adds depth information to Bag of Visual Words' feature extraction part which is a novel approach in the literature. The proposed 3D descriptor eliminates the disadvantages of brightness-based problems and improves the structure with depth information. This improvement gives better accuracy results compared to the original method providing a rational and useful method for 3D object detection.