Research On Segmentation Based On Superpixels For Family Environment

Indoor robots are currently the closest robot to human life.They can replace or assist people to complete some affairs in life,and gradually become an indispensable part of the family.As the eyes of robot,vision system which is an important part of robot intelligence has been able to perform the tasks such as image classification,detection,and segmentation As the basis of scene understanding,semantic segmentation aims to segment the scene into several areas consistent with human visual perception,so that the robot can recognize the typical objects in the environment,and play a key.role in robot intelligence.However,due to the complexity of the indoor scene,the existing semantic segmentation model has the problem with inaccurate recognition.The main manifestations are sensitive to light and the weak recognition of low-level features of the image,resulting in inaccurate image edge,which affects the subsequent related tasks for robots.Therefore,good algorithm performance is of significance to intelligent application of robots indoors.Aiming at the inaccurate segmentation of FCN for indoor scene,the paper makes some new attempts on the basis of this model,and main contents are summarize as follows:(1)For the problem of weak recognition of low-level features in semantic segmentation,we propose to improve the ability of semantic segmentation model by using superpixel algorithm based on the research of related theory and technology of image segmentation After studying and comparing several typical superpixel algorithms,we analyze the feasibility of SLIC algorithm for improving the effect of semantic segmentation.(2)For the problem of the SLIC algorithm due to the sensitivity to image brightness and the limitation of superpixels,we propose a method by introducing HSL color space and Sobel operator.Based on SLIC algorithm,the method improves the result of superpixel segmentation by using three independent components of HSL color space and 8-connected regions of Sobel operator,which improves the robustness of the SLIC algorithm and makes segmentation more adhered to the image edge.(3)For the problem of FCN model due to the light sensitivity and weak recognition of low-level features,we proposes a semantic segmentation method based on superpixel algorithm.The method uses superpixel algorithm to perform local edge processing on the rough segmentation results obtained by the FCN model,and then uses a fully connected conditional random field for overall optimization.The accuracy of this method is better than the existing algorithms,and it can effectively segment the target object indoors.(4)B-uilding an intelligent grasping platform for indoor robot,we combine the proposed semantic segmentation model and the method of pose estimation to recognize and grasp the target object.The proposed method will be verify by the experiments of robot grasping in the real scene.