Research On Siamese Network Tracking Algorithm Based On Spatial And Frequency Domain Features

Target tracking is an important research topic in the field of computer vision.With the continuous development of deep learning technology and large-scale test data sets OTB,VOT,Got-10 k,etc,target tracking algorithms based on neural networks have received extensive attention.The Siamese network series is an important branch of the single target tracking algorithm.By describing the target tracking problem as learning a general similarity mapping function problem,the tracking accuracy and time complexity are well balanced.It follows the Siam RPN,Siam RPN++ and other algorithms.The emergence of,got rapid prosperity.The subject summarized the twin network tracking algorithms into a unified target tracking system model,and focused on the analysis of the key feature extraction modules.Different from other related studies,the subject does not focus on the network structure and algorithm design,but separates the spatial and frequency domain features from the perspective of feature extraction,and designs the corresponding algorithm frameworks respectively.For the traditional airspace feature network,previous research mainly focused on introducing a deeper and more advanced backbone network for the twin network target tracker,while the current paper focuses on improving and upgrading the network architecture,and striving for the deep and shallow features.The balance and supplement.In view of the difficulty of combining deep and shallow network structures,as well as complex engineering problems such as redesigning algorithms and retraining networks,consider introducing simple and easy-to-implement traditional appearance information,and the description of the target will be more intuitive and specific.The subject designed an enhanced airspace algorithm based on apparent information to further enhance the degree of discrimination of apparent information for targets within the class.The framework adds three modules to the Siam RPN++ algorithm structure: apparent features,voting candidate mechanism and adaptive search area module.The voting candidate mechanism is to facilitate the introduction of the apparent feature module,while the adaptive search area strategy is to improve the algorithm framework from the perspective of model update.After testing,the scores of the designed algorithm framework on the VOT data set are all in the forefront of the challenges over the years,which greatly improves the performance of the original algorithm.Secondly,the complex and diverse image features are not only manifested in the spatial domain,but the frequency domain features are also a very important feature.Aiming at the current deep target tracking algorithm focusing on the spatial domain and using the spatial features of CNN,the subject considers the introduction of frequency domain learning for the first time in the twin network series trackers.The filter under the frequency domain feature is better than the downsampling operation under the spatial feature.This makes it possible to use the frequency domain algorithm framework to better retain the original accuracy of the algorithm if it is reduced to the same feature dimension to increase the speed of the algorithm..Aiming at the frequency domain algorithm framework,the subject adopts the Siam FC algorithm,which has the common characteristics of twin network tracking algorithms and is easy to be extended later,as a benchmark.It uses DCT transformation,adds frequency domain preprocessing and static channel selection operations,and finally tests under the OTB data set and scores the results The feasibility of the frequency domain algorithm framework is verified.At the same time,whether it is forward inference or reverse training process,the frequency domain tracking algorithm with feature trimming has certain computational performance advantages.