RGB-D Image Instance Segmentation Based On Seeded Region Growing Algorithm

Intelligent gripping and handling robots can efficiently and reliably complete various tasks to significantly reduce labor cost. Among other things, accurate object detection and location forms the basis for the robots to function properly. In this paper, a novel target detection framework based on RGB-D images in an unstructured environment has been proposed. First, we use the instance segmentation network Mask R-CNN training model to pre-segment the target object and produce rough initial masks under the condition of small sample training. Then we apply the seeded region growing algorithm to automatically generating seed points. In the proposed method, we incorporate the depth information into the merging principle of the region growing algorithm to leverage three-dimensional information of target objects. This improves the accuracy and robustness of the object segmentation considerably. Experiments show that the segmentation results under proposed method contain the instance information of the target while maintaining high level of accuracy. This is crucial for downstream activities, i.e., the subsequent robotic grasping.