Research On Operation Object Recognition And Localization Of Cooperative Robot Based On Deep Learning

The problem of object recognition and localization is an important branch of collaborative robot research.The method of object recognition and localization is the key technology for collaborative robots to complete tasks such as handling,sorting,palletizing,assembly,and quality inspection.At present,object detection algorithms based on deep learning can realize robot operation object recognition,and have high accuracy,which is a research hotspot in the field of robotics.Based on the characteristics of the object in the collaborative robot operation scenario,research on the method of recognition and localization of the collaborative robot operation object based on deep learning to improve the detection accuracy has important theoretical and practical significance,and also has broad application prospects.In this paper,the research object of cooperative robot operation object recognition and research is carried out.The research of object detection algorithm based on deep learning and the research of object positioning method based on binocular camera are carried out.The main research work of the paper is as follows:First,the method of object detection based on feature fusion is studied.Aiming at the problem that the deep learning object detection algorithm SSD(Single Shot MultiBox Detector)based on convolutional neural network has a poor recognition effect on small scale objects,an extended deconvolution SSD object detection model based on multi-scale feature fusion is proposed.The model first selects a specific layer of the basic network,and expands features by upsampling and summing elements to improve the representation of the feature pyramid.Then,based on the new feature map obtained from the expansion,a multi-scale feature pyramid is constructed to achieve multiple Scale object recognition;Finally,the shallow feature layers of the feature pyramid are selected to perform feature fusion by deconvolution upsampling and feature maps in series,enhancing the semantic information of the shallow feature layers,and improving the detection effect of small object objects.Test results on the public data set of PASCAL VOC show that the average accuracy of small object objects is improved.Secondly,the method of object detection under occlusion is studied.Aiming at the problem that the non-maximum suppression algorithm in the deep learning object detection model deletes the occluded object detection frame by mistake,a continuous softening non-maximum suppression algorithm is proposed.The classification confidence score reset function in the original non-maximum suppression algorithm was changed from the zero setting function to a continuous attenuation function,which not only ensured the suppression effect of the detection area with excessive overlapping area,but also measured the remaining The scores of the boxes are attenuated to different degrees,and there is no sudden mutation in the confidence scores of the remaining detection boxes.This reduces the probability that the obstructed object is judged as a false positive.The validation on the public data set of PASCAL VOC shows that the recognition accuracy is improved in the case of occlusion and has no effect on the case of no occlusion,which proves the effectiveness of the improved method.Finally,the method of object positioning for collaborative robots is studied.The method of anchor frame setting in the deep learning object detection model is deeply researched,and the influence of the size,proportion and quantity of anchor frame on the detection effect is analyzed experimentally.According to the actual scene,a collaborative robot operation object data set is produced.The clustering method is used to extract the object scale and scale features in the dataset.The anchor frame scale and scale are recombined with this feature and applied to the improved SSD object detection model.The detection results prove that the accuracy of the model detection after the modification of the anchor frame setting has been improved to a certain extent,and the object in the image can be recognized.According to the result of object detection,the position of the cooperative robot operation object in the image can be obtained.In combination with the binocular camera,the spatial positioning of the operation object is achieved.