IJCAS

Deep learning models are used to track target-of-interest (ToI) objects with autonomous mobile robots (AMRs). A Siamese network based on few-shot learning is used when there is limited training data for the ToI objects. A Siamese network recognizes the object’s class based on the similarity between the input object and the template of the ToI object. The more similar the shape of the objects to be recognized, the smaller is the interclass variance of the features extracted from the object and the larger is the intra-class variance. If the inter-class variance is small and the intra-class variance is large, it is a different from the template object; however, since the calculated similarity can be high, it can be recognized as the same class object. Alternatively, it is an object of the same class as the template object, but can be recognized as a different class of object. To increase the inter-class variance and decrease the intra-class variance, a Siamese network training method is required. This paper proposes a method that recognizes ToI objects by increasing the difference in features between the ToI object and non-ToI objects when recognizing an object using a Siamese network. When distinguishing the ToI object from non-ToI objects, images that act as an anchor for each class are selected. The images inputted to the Siamese network are trained such that the corresponding class is close to the anchor of the selected class and far from other classes. The trained Siamese network then compares the anchor and similarity of each class to determine whether it is the ToI object. An experiment was conducted to analyze the training and execution results of the Siamese network using the proposed method. The recognition results using the proposed method and an existing template-based Siamese network were compared. The convolutional neural network model trained with the proposed method yielded an average accuracy of 93.95%. The proposed method could improve the performance by 23.68% compared to the existing template-based Siamese network.

Keywords: Convolutional neural network, deep learning, few-shot learning, mobile robot, object recognition, Siamese network.