Research On Object Detection Of Desktop Robotic Arm Based On Background Subtraction And Few-Shot Learning Classification

With the reduction of hardware costs,robotic arms widely used in the industry have entered people's daily lives in the form of low-cost and of lightweight desktop robotic arms.Compared with a single task of the industrial robotic arm,the desktop robotic arm requires a more flexible object detection method to handle various tasks in daily lives.To the best of our knowledge,there exists two object detection methods.One is traditional object detection methods,which need additional features to descriptors of the object in a model.Another is object detection methods,which based on deep learning need use a number of object data to train the neural network model.If the model is not modified,then the above two methods cannot be migrated to the detection of new categories.That is,their models are highly dependent on the object to be detected.However,a number of desktop robotic arm users without professional knowledge can only use the object detection system provided by the robotic arm when it leaves the factory.This system cannot detect objects that are not present in the system.Moreover,users cannot change the objects operated by the robotic arm according to their own needs.Because the few-shot learning method can directly transfer the learned general knowledge to the processing of a new task,the object detection method based on fewshot learning can solve the problem that the model is highly dependent on the object to be detected.However,the existing object detection methods based on few-shot learning have low accuracy,mainly due to the following shortcomings:(1)Under the premise of a small number of samples,neither the method based on sliding window,region proposal nor the method based on anchor can accurately locate the object.These methods not only require regression processing for positioning,but also generate a large amount of useless background,which increases the difficulty of classification of the model.(2)The existing public data sets have the characteristics of unobvious objects,complex backgrounds and large differences between the same types.This requires that the model must have a good generalization ability to obtain satisfactory results.However,the few-shot learning is contrary to the generalization ability.In this thesis,considering that the camera and background are both static in the desktop robotic arm scene,and taking advantage of the good transferability of few-shot learning.This thesis proposes a desktop robotic arm object detection method based on background subtraction and few-shot learning classification.Regarding the above shortcoming(1),this method does not use sliding windows,region proposal and anchor to locate the object,but uses a foreground extraction method based on improved background subtraction to accurately locate the object and perform pixel-level segmentation of the object,and there is no need to regress the positioning.Then directly use the few-shot learning classification method based on the improved Relation Network to classify the segmented objects.Since the object obtained by segmentation is a significant object that does not contain the background,the classification result has a very high accuracy rate.Regarding the above shortcoming(2),this method uses the foreground extraction method based on improved background subtraction to extract images with no background and significant objects from video frames,and then uses these images as a data set to train the few-shot learning classification model based on improved Relation Network.This operation further improves the accuracy of classification.Specifically,the work of this thesis is as follows:(1)This thesis proposes a foreground extraction method based on improved background subtraction.(1)By installing an auxiliary board,this method solves the problem that the camera is prone to shift in the process of interacting with the user,and can also eliminate the interference of non-object foreground such as human hands.(2)For the problem that multiple objects overlap and cause the object foreground to be connected,this method uses the segmentation method based on the watershed algorithm to segment each object foreground.(3)For the problem of a lot of noise in the multi-texture background environment,this method proposes a three-frame threshold method.(4)For the problem that it is difficult to extract the object foreground similar to the background,this method proposes the difference degree to describe the pixels.Experiments show that this method can work normally when the light changes,the object foreground is similar to the background,and the camera or background is slightly shifted.(2)Made 5 improvements to the few-shot learning classification model based on improved Relation Network.(1)The essence of the Relation Network is to measure the similarity between Support Set and Query Set,which is a regression problem.However,the existing public data sets only have category labels,and regression processing cannot be performed on these data sets,so this method uses a weighted color histogram similarity measure to label the data sets for similarity.(2)Considering that in desktop robotic arm applications,there is no need to distinguish between the same types of objects that are too different,and the opposition between few-shot learning and generalization capabilities.Therefore,this method changes the question of "which type" of the few-shot learning classification into a question of "which one",and making the question more suitable for the similarity measurement problem.(3)Analyzed the reason for the high accuracy of the few-shot learning classification method on the dataset Omniglot,but the low accuracy on the data set mini Image Net,and established the data set NBI(No Background Image)that is more suitable for few-shot learning classification.(4)For the existing methods to fix the N and K values in the N-way K-shot task when training the network model,the training method will cause the model to depend on the N and K values.The Rand-way Rand-shot training method is proposed.(5)The color histogram similarity measurement module is added to the model to measure the similarity of the shallow features between samples,which is used to correct the obvious misjudgment of the Relation Network.Experiments show that these improvements have significantly improved the accuracy of the few-shot learning classification method based on the Relation Network.(3)Based on the result of positioning the object in the pixel by the above method,a method of positioning the object in space without hand-eye calibration is proposed.Experiments show that the object detection method can perform accurate object detection in different desktop robotic arm applications and give the object's twodimensional position and posture.Moreover,when adding a new object category,only a few photos of the object of the new category need to be taken as the Query Set to realize the detection of the new category,and non-professionals can also perform the operation.