| The enhancement of the robot’s ability to perceive the external environment is very important for it to realize various fine operations.As two modes of direct interaction between robot and external environment,visual and tactile play a direct role in improving robot perception ability.Vision can provide appearance information such as shape and color of an object,while touch can provide attribute information such as texture,roughness and hardness of an object.The integration of visual and tactile information can make up for the limitations of single perception mode and provide more complementary information for object classification.However,fast and accurate classification of objects in massive databases is more important in practical production and life.Therefore,this thesis focuses on how to further improve the efficiency and accuracy of object classification and carries out the following work:(1)In order to extract the feature information of different levels of tactile signals,this thesis proposes a Tactile Net with four-layer one-dimensional convolution as the basic network of tactile feature extraction,and through comparative analysis of several visual feature extraction models,Alex Net is selected as the basic network for the extraction of visual feature.Due to the limited information contained in a single capture of tactile signals,it is found in this thesis that the model cannot accurately judge the changes of tactile signals.In view of this,this thesis divides every two grasping processes of tactile data into an independent tactile signal.The experimental results show that the processing method proposed in this thesis makes it easier for the model to capture subtle changes in tactile signals,and thus improving the accuracy of model classification.(2)In order to fully verify the rationality of the method in this thesis,the visualtactile adjective dataset VHAC-52 is developed,which contains the visual and tactile data of 52 kinds of common objects and 14 adjectives used to describe the objects’ visual and tactile attributes.Compared with the PHAC-2 and VHAC datasets,this dataset provides richer and more comprehensive information about the objects,laying the foundation for more accurate object recognition and more adequate model training and testing.(3)The direct fusion of visual and tactile feature vectors will lead to redundant feature information,which makes it difficult to improve the speed and accuracy of model classification.Therefore,this thesis proposes a visual-tactile fusion object classification method based on adaptive feature weighting.In this method,the channel attention mechanism is used to adaptively calibrate the feature responses between channels,which can enhance the important information and weaken the background,noise and other minor information,so as to improve the sensitivity of the network to the discriminant feature information,and thus effectively improving the classification ability of the model.The experimental results demonstrate that for the public dataset PHAC-2,the classification accuracy of the proposed method is 0.9750 on the AUC(Area Under Curve),which is 1.92% higher than that of the known best method.For the novel VHAC-52 dataset,the proposed method achieves values of 0.9827 and 0.9850 in terms of the AUC and ACC(Accuracy)metrics,respectively.Furthermore,the inference time reaches 1.559s/sheet,demonstrating the superiority of the proposed method. |
