Research And Application Of Indoor Object Detection Based On Hand-held Object Learning

In indoor scene,intelligent robots often need to perceive and understand the scene,in which object detection plays an important role in human-computer interaction,navigation and path planning,motion control etc.Commonly used methods of object detection utilize many labeled data and are applied in indoor scene.However,in real scenarios,the robots need to learn object concept information in the process of interaction,and transfer it to detect objects in indoor scenes.Therefore,based on handheld object learning,this paper proposes a framework of object detection in RGB-D indoor scene,so as to realize the task of object learning and detection in humanmachine interaction.In the hand-held object learning process,intelligent robots need to segment the object in hands.During object detection in indoor scene,intelligent robots also need to get object proposals through segmentation.Due to the distribution diversity between hand-held object data and object data in indoor scene,in this paper,deep unsupervised adaptive network,a deep neural network based on adversarial learning,is derived to adapt feature matching,which reduces the deviation between two different domains.The main work of this paper is as follows:(1)An improved region growing algorithm,based on human skeletal data and depth image information,is proposed for hand-held object segmentation in the human-computer interaction process.(2)To gain object proposals in indoor scene,random sampling consistency(RANSAC)algorithm and Euclidean distance clustering method are used to fit the plane and cluster object point-clusters.After that,the segmented regions in color image and depth image are obtained by the camera parameter estimation.(3)Aiming at the problem that the hand-held object data and the actual indoor scene data are obviously different and the distribution is inconsistent,based on the thought of deep unsupervised adaptive neural network,the indoor object detection framework based on hand-held object learning in RGB-D indoor environment is proposed,which makes the hand-held objects match theobjects in indoor scenes of the target domain.Through the experimental analysis,where training and testing with image with both RGB and depth modalities,the detection accuracy of the deep unsupervised adaptive neural network is higher than that of the common deep neural network.This suggests that the technique of transfer learning used in deep neural networks can make hand-held object data more adaptive to object data in indoor scenes,and improves the performance of object detection in interactive environments.