The Research And Implementation Of High-Speed Tracking Algorithm Via Convolution Feature

In recent years,Convolution neural network has gained remarkable achievements in many computer vision tasks due to its powerful feature extraction ability.However,the application of convolution neural network in visual target tracking field has been limited,because convolution neural network with strong feature extraction ability is deep.And the cumbersome computation process of convolution feature extraction leads to tracking speed performance suffer greatly affected,far from satisfying the needs of high-speed target tracking in reality.Most of the tracking algorithms use convolution neural networks trained by other visual tasks to extract convolution features,and combine with correlation filtering framework's fast detection ability to accomplish tracking tasks.These algorithms can achieve good tracking accuracy,but can't satisfy the requirement in high-speed target tracking.This thesis will discuss the way of combining convolutional neural network and correlation filter framework,extending the potential of convolution neural network in highspeed target tracking.The contributions of this article are mainly as followings:(1)Combine Siamese network framework with correlation filter framework to realize a tracking algorithm.Using convolution neural network trained from other visual tasks to extract convolution feature,then utilizing correlation filter framework for fast detection and tracking.The above method separates the association between convolution feature and correlation filter framework.In our paper,correlation filtering framework is added to offline training process of convolution network.In off-line training process,the spatialtemporal structural characteristics of the correlation filtering framework is exploited.After training,the convolution network can extract convolutional features including spatialtemporal information.Besides,the fundamental convolution network structure adopts Siamese network framework,which has strong feature extraction capability.Our tracking algorithm coalescing Siamese network framework with correlation filter framework is named Siam CFnet.The OTB2013 is used to evaluate Siam CFnet algorithm implemented in our paper.Experimental results show that Siam CFnet algorithm can extract discriminative convolution features only by using shallow convolution network,achieving better tracking performance and keeping 80 FPS tracking speed.(2)To overcome the shortcomings of Siam CFnet algorithm,a tracking algorithm of fusing improved Siamese network framework and correlation filter framework,named Siam CFnet_new algorithm,is proposed.In order to further enhance the feature extraction ability of convolution network in Siamese network framework,residual module and channel attention mechanism are proposed to improve the structure of convolution network and enhance the appearance representation ability of convolution features.The correlation filter framework is optimized by appending scale estimation mechanism and the strategy of template updated adaptively,making our tracking algorithm more robust in background clutters and occlusion.The evaluation results using OTB2013 show that Siam CFnet algorithm after optimizing Siamese network framework improves the success rate by 6.48% and the accuracy by 8.01%.The tracking performance of Siam CFnet algorithm after optimizing correlation filtering framework has also been significantly improved.Finally,the Siam CFnet_new algorithm proposed in our paper are evaluated and compared by using OTB2013 and OTB2015 evaluation sets.The results show that compared with Siam CFnet algorithm,Siam CFnet_new algorithm improves the success rate by 16.05% and the accuracy by 15.35%,and get no weaker performance than min current excellent tracking algorithms.(3)Design GPU's acceleration project for multi-scale Siam CFnet algorithm.The parallelism of multi-scale Siam CFnet algorithm is analyzed,and the parallel design of each module of the algorithm is introduced.Using CUDA architecture to achieve acceleration and taking speed's evaluation.The experimental results show that the speeded multiscale Siam CFnet algorithm can achieve minimum speed of 200 FPS and maximum speedup of 30.It has high precision tracking effect and can be applied to high precision high-speed target tracking scenarios.