Improving 3D Object Detection and Classification Based on Kinect Sensor and Hough Transform

Hough Transform has been successfully applied to a variety of image processing problems in recent years. This papers presents a novel approach for detecting and classifying 3D objects by using the generalized Hough method and the Kinect (TM) sensor. Our algorithm considers feature points and color spectra as two interleaved processes to cooperatively recognize objects in a 2.5D fashion. With this strategy, the algorithm automates the image pre-processing operations regardless of scenes (i.e., particle cleaning, hole filling, particle eroding, and object dilating) and reduces the processing load over the sensor's point cloud for 3D object classification. Extensive experiments applied but not limited to recognition between different and similar objects, occlusion, and perspective change analyzing fitness and processing time show that the 2.5D approach makes feasible 3D object recognition for applications with video information.