Research On Siamese Network Object Tracking Algorithm Based On Migration Learning

Object tracking is a fundamental research element in computer vision,which refers to the tracking of an object in a video or image sequence in order to continuously obtain the object’s motion parameters and gain an understanding of the object’s behavior.With the public availability of large annotated datasets and improvements in the processing power of computer hardware,much progress has been made in related research.However,object tracking is still a challenging task due to the diversity of the actual environment in which the object to be tracked is located and its complex and irregular movements.In addition,traditional neural network training methods are difficult to achieve satisfactory results when faced with object-specific tracking tasks or when tracking objects captured from the UAV’s perspective.In this thesis,an in-depth study is carried out based on the analysis of existing tracking algorithms.The specific work is as follows:(1)A Siamese network object-tracking algorithm based on migration learning is proposed.Firstly,we design an object-tracking algorithm based on the region candidate network(RPN),select the Alex Net network for feature extraction in the backbone part,and pre-train the network using typical datasets;secondly,we manufacture multi-rotor UAV sets,including training and testing.Then,according to the pre-training strategy,some parameters of the neural network are frozen;the remaining parameters are fine-tuned by the training set,and the performance metrics of the tracking drone are obtained by the testing set from the experiment.Finally,a quantitative measure of the internal features of the deep neural network is proposed,the similarity score formula for deep attribution mapping,which is applied to the inner characteristics of a layer of the Alex Net network in the UAV tracking algorithm.The experimental results verify the effectiveness of the proposed method.(2)A migration learning-based object tracking algorithm for temporal Transformer networks is proposed.Firstly,the Alex Net network is improved by using temporal convolution for feature extraction,and the temporal information of the object is introduced from the backbone part of the tracking network;secondly,the feature map containing temporal information is adaptively weighted using the channel attention mechanism to make it more focused on the information of the current frame;finally,when establishing the information interaction between the template frame and the search frame features,the Transformer module for feature fusion is introduced to obtain the temporal information of successive frames during the tracking process.After training using the migration learning method,testing is performed on 4 authoritative standard UAV datasets and 3 custom datasets.The experimental results validate the effectiveness of the proposed algorithm and the applicability of the migration learning law summarised in the first point of work;and the algorithm runs at a frame rate of 12 FPS when deployed on the NVIDIA Jet-son TX2,providing a reference for the deployment of the algorithm in practice.