Research On Robotic Perception Of Object For Intelligent Grasping

As the development of intelligent technology,robots have already been applied to various fields,including industrial manufacturing,education,entertainment and household service.However,the robotic intelligence is still quite limited.Despite these robots are with abilities of obstacle avoidance,speech recognition and command feedback,these abilities are quite fundamental and generally valid only in simple scenarios.Assisting and interacting with human in a higher and more intelligent manner is still a challenging task,where robotic grasping has become the key and promising technique to achieve this.Unfortunately,the robotic sensing ability to object as well as the environment is relatively weak,which suffers the challenges of environmental complexity and uncertainty,the diversity in object category,the occlusion between objects and cluttered background.This work takes grasping object under unstructured scenarios as the background,focuses on the robot equipped with multi-DOF rigid arm and studies the challenges in its visual perception.The main work is summarized as follows.Considering the template-based methods could not be directly used to detect objects with unknown categories,we propose a class-agnostic object detection method using the region proposals.It uses the under-sampled histograms of oriented gradients to distinguish objects from background.Meanwhile,the pre-processing strategies,such as local normalization and wiener filtering,are integrated to enhance the contrast between foreground and background.Experimental results show that the proposed method could detect possible objects in the image with efficiency and is robust to rotation,imaging view,illumination and noise.As for the border bias between predicted region and ground truth region,we propose an object detection refinement method based on local features,which compensates the bias from the aspects of visual saliency and edge distribution respectively.Considering the feature consistency in object,the three-stage ”Expansion-Alignment-Optimization” regional border optimization method is proposed.Specifically,the superpixel-based sliding window strategy as well as the foreground-background classification method based on high-contrast superpixels are adopted to optimize the border position of the predicted region.The effectiveness of the proposed method is verified on both the public dataset and real-world scenario.Experimental results show that our proposed method could improve the detection rates of other proposal region generation methods,while keeping their algorithmic frameworks unchanged.Considering the high computational burden for traditional grasp search methods,a selective search based grasp detection method is proposed.In graspable regions generation,we design two gradient-based grasp criteria(friction cone and grasp torque)to select the feasible pairs of grasp points and determine the parameters of grasp rectangle according their Euclidean distances.It could achieve the 99% overlap with ground truth using limited number of grasp candidates.Then in grasp rectangle recognition,a convolutional neural network is constructed to recognize the graspable regions,which only takes the RGB image as input.The comparative experiments are deployed to prove that the proposed method could achieve competitive performance with traditional deep learning based methods and greatly reduce the algorithmic training time.Considering that the limited number of ground truth grasps are difficult to fully reveal the grasping probability of the predicted grasp,we propose the grasp detection method using the grasp distribution line.It could describe the consecutive distribution of valid grasps.Through comparing the predicted region with the mapped region,instead of the ground truth one,a fairer and more accurate measurement of the model loss could be achieved.Besides,we further propose the multiple grasp detection method using YOLO v3 as the backbone,which divides the problem of grasp detection into the regression of its location and size,and the classification of its orientation.It could accurately detect multiple grasps on the object and achieves the detection rate of 96.5% in Cornell Grasping Dataset.When applied to the real-world robotic grasping scenario,it could successfully guide the rigid arm to accomplish the task of grasping object.