Transparent object detection using color image and laser reflectance image for mobile manipulator

It is a significant challenge to detect a transparent object using conventional visual sensors. In this paper, we show that the characteristics of transparent objects under different lighting conditions (normal light and laser light) can be exploited for improved detection accuracy. The authors analyzed the object characteristics and found a novel method for transparent object detection. GrabCut[11] method is used for the image segmentation. However, the transparent object cannot be detected from color or depth images using GrabCut, because transparent objects always shows the color of background. So, we propose a method of transparent object detection using not only color image but also laser reflectance intensity image. Highlight spots from color image can be used for detection candidate areas of transparent objects. These candidate areas is used for GrabCut as bounding rectangle input. Thereafter, segmentation by GrabCut is applied to depth images and laser reflectance intensity images. Bottle-shaped transparent objects are correctly detected within short distance by the visual sensor using the proposed method. In addition, other evaluations show the robustness and the limits of our method for different shapes, materials, distance and background.