Research On Real-Time Single Object Tracking Algorithms

Single object tracking is one of the fundamental problems in computer vision.In single object tracking,after determining the object in the first frame,one needs to give an approach,which enables the computer to track the object accurately and continuously in the subsequent frames.Single object tracking has a wide range of applications in many domains,e.g.intelligent video surveillance,motion recognition and so on.However,single object tracking is still a challenging task due to the influence of complex factors such as real-time,fast motion and illumination variation.This paper focuses on the real-time single object tracking algorithms.The main contributions are as follows:(1)A shallow learning algorithm FSCF(Fuzzy Least Squares Support Vector Machine based Correlation Filter)based on multi-feature fusion is proposed.In the existing algorithms based on bilateral weighted least squares fuzzy support vector machine,there are some problems such as high computational complexity and sensitivity to deformation.In order to solve these two problems,we propose a real-time tracking algorithm FSCF based on multi-feature fusion.In the proposed algorithm,for the local HOG(Histograms of Oriented Gradients)features,the correlation filtering framework is used to overcome the time-consuming matrix inversion operation,while the training data and background information are augmented through multibase samples to alleviate deformation-sensitive problem.For the global color features based classifier,one-hot encoding is used to achieve fast computation.Finally,we use the linear combination of the two classifiers to realize object tracking.(2)A deep learning algorithm Siam FC?plus(Siamese Fully Convolutional Plus)based on 2-channel network is proposed.In the existing algorithms based on siamese network which is represented by Siam FC(Siamese Fully Convolutional),how to learn the similarity function effectively in the offline training phase is the bottleneck of the performance improvement.Considering the accuracy and the efficiency of the 2-channel network in metric learning,we introduce the 2-channel network into the single object tracking field,and propose a real-time tracking algorithm Siam FC?plus.In the proposed algorithm,firstly,the exemplar image and the search image are fused together through cross-correlation operation.Then the deep features are extracted through the convolutional neural network.After that,the decision layer outputs a response map.Finally,the position with the highest response value is set as the new position of the object.(3)In order to verify the effectiveness and robustness of the two proposed algorithms,extensive experiments are carried out on the benchmark datasets.The results demonstrate that our algorithms can achieve comparable or superior performance to many state-of-the-art algorithms.