3d Poind Cloud Based Active Object Recognition System

In recent years,it has been a new trend to equip a intelligent robot with inexpensive RGBD cameras(e.g.Kinect)as its visual perception system.Object recognition is one of the most important functions of its visual system,and has always been a challenging topic.The main reasons lie in the uncertainties in a real environment(table-top environment where objects are cluttered): In a single view,when objects are cluttered on a table,the lose of information caused by occlusion will sharply reduce the accuracy of recognition and pose estimation.Due to this limitation,researchers have proposed active object recognition systems for mobile robots,that is to say,they solve the occlusion problem by observing from multiple views.However,in practical robots have to move in a large area and the choice of view points are usually constrainted by environment.To this issue,in this thesis we propose an active object recognition system based on the robot-environment interaction.Its main idea is to use robot arms to rearrange the cluttered objects,in order to actively search and recognize objects.In detail,determine the occlusion relations between object hypotheses though scene analysis and recognize the objects to be moved.Then use robot arms to change their locations and poses,and increase the visable surface of occluded objects and further promote the recognition rate.In this article,we focus on scene analysis,and our contribution is described in the following three parts:1.We propose a 3D point cloud scene segmentation method.First,the background(e.g.table)points are filtered from the input point cloud.Then the rest cloud is over-segmented into supervoxels and they are treated as the basic elements of our segmentation method.The third step is to merge supervoxels into surface patches according to their normals.The surface patches are further merged into object hypotheses based on two physical characters of a normal object surface using graph-cut method.Finally,object borders,separated object parts and open containers are optimized.We use benchmark dataset to verify our method and our method performs as well as the state-of-the-art methods.2.We propose a method to determine the occlusion relations among the segmented object hypotheses in a range image.It aims at sorting the segmented object hypotheses into two categories:occluding objects with complete surface information,occluded objects with uncomplete information.To do so,the method searches object border and shadow point in a range image.In our system,we only recognize and move the occluding objects.3.We propose a object recognition and pose estimation method which combines both global and local features.Given a complete object surface,the advantage of using a global feature is its low computational cost,but the weakness lies in its unaccurate pose estimation;in contrast to that,pose can be better estimated using local feature in the price of longer time.Therefore,we combine these two advantages and proposed a novel recognition method.We also verify this method in a benchmark dataset.Finally,we conduct experiments in real scenes to verify the proposed active recognition system.We build a household environment including generating 3D models for several objects.The experiments prove that our recognition system increases the cognitive ability and intelligence of robots.