Three dimensional object segmentation based on spatial adaptive projection for solid waste

Automatically waste sorting is usually resorted to the robot arm to grasp solid waste. However, the solid waste is prone to be stacked up or connected, and they are difficult to segment and consequently to be grasped. To resolve this problem, we present a novel spatial adaptive projection method based on only one RGB-D sensor to accurately segment the piled or connected solid waste. After obtaining the point cloud of waste objects from waste disposal line through the RGB-D sensor, a 3D background-difference is used to remove the background, and Euclidean Cluster Extraction is employed to obtain a set of point clouds. Because most of the waste are stacked up or connected, each point cloud may contain multiple objects. For each point cloud, we deal with it by some preprocessing, then project it into different planes to get an optimized image which is easy to be segmented. The projection parameters are automatically computed based on a mathematical optimization framework. After segmenting waste objects into several isolated objects, we re-project them into the point cloud. Before generating the final segmentation, a check mechanism that avoids over-segmentation would be used to judge the relationship among these objects. In addition, we established a point cloud dataset of waste objects, which contains stacked up and connected cases, and we evaluate the effectiveness and accuracy of our method in this dataset, the results show the performance of the proposed method is satisfactory.