Few-shot Learning Method Based On Multiple Distance Metrics And Its Application In Target Tracking Task

Deep learning plays an important role in image recognition and other visual fields because of its powerful capabilities of feature extracting and learning.However,the conventional deep learning methods rely on a large amount of labeled data for training,and it is difficult to exert its value when lacking of labeled samples such as target tracking task.In this paper,we analyze and design a few shot learning algorithms and applies it to target tracking task.The main work of this paper is as follows.(1)We transform the target tracking task into a few shot learning task according to the characteristics of the target tracking task that lacks labeled data.And we improve the few shot learning algorithm for the interference caused by the complex background encountered in the tracking task.We propose a network structure of multi-distance measurement by making full use of the shallow features of the image.The result can be judged based on the local features in the shallow information In the case of small differences between different categories.(2)We also propose an adaptive inter-class distance loss function method according to the characteristic of different environmental factors faced by different tracking tasks.We train the model through a metric learning method and introduce triplet loss.What is more,we extend the distance loss to few shot learning training tasks and increase adaptive loss calculations to enhance the model's anti-interference ability and improve the model's adaptability to different tracking tasks.(3)In this paper,we successfully applied the improved few shot learning algorithm to the target tracking task.Besides we propose an end-to-end target tracking model based on multiple distance metrics method.We design an adaptive template update strategy based on metric learning by considering that the conventional methods only relying a unique template and not conducive to long-term tracking tasks.So that the frequency of template update can be dynamically adjusted according to the target's changes in different time periods,and enhance the robustness of the model.Based on the similarity between target tracking task and few shot learning task,we combines the two and propose a target tracking framework based on few shot learning,and achieves ideal experimental results on the standard data set.It proves the few shot learning method proposed in this paper and its application in target tracking tasks are effective and have practical significance.