Research On Correlation Filter Tracking Based On Deep Feature

Visual object tracking,as an important research field in computer vision,has been widely used in video surveillance,human-computer interaction,intelligent transportation and navigation,and military target positioning.While great progress has been achieved in the past 60 years,the visual object tracking performance is limited due to the influence of many factors such as the deformation of object appearance,occlusion,illumination variation and background clutters,and there is not a common tracking algorithm that can solve all the above problems.It is a challenging task to design an object tracking algorithm with strong robustness,high tracking precision and good real-time performance.In recent years,the correlation filter theory based the tracking methods have achieved superior performance on the visual object tracking dataset.In these approaches,a discriminative correlation filter(DCF)is trained firstly during the initialization phase,with subsequent steps alternating between detection and update.1)Detection: the object appearance in the current frame is obtained from the position in the previous frame,and the correlation response map is getted by doing correlation operations with correlation filter.The object position in the current frame is determined by the peak of the response.2)Update: we first extract the object appearance in the current frame using the new object position,and then update the correlation filter according to the expected output.As the above operations can be implemented in the frequency domain by the Fast Fourier Transform,the correlation filter based tracking has a very significant real-time advantage.Based on this framework,this paper mainly focused on the improvement of the appearance model,training and updating strategies,as well as dealing with scale variation.Convolutional Neural Networks(CNNs)have emerged in recent years with powerful image feature representation.In this paper,we combine the tracking algorithm based on correlation filter and deep neural network to study the object appearance and scale adaptation strategy.The main contents are as follows:(1)On the basis of extending the correlation filter from single channel feature to multi-channel,the correlation filters are trained for object tracking by different object appearances representation,including,the original gray features,HOG features,the deep features from different convolutional layers in CNNs.By exploiting the transfer learning characteristics of VGG network,the output in each layers are extracted from the VGG network model imagenet-vgg-2048-network which is designed for image classification task in the MatConvNet library,is taken as a feature.The above features were applied to the tracker and tested on the OTB-2013 dataset.The experimental results demonstrate that the deep features have a significant advantage compared to conventional hand-crafted ones,and the best results are obtained using features in the first and fifth layers.Based on the above observation,we propose to train the correlation filters separately in the first layer and the fifth layer,and localize the target precisely by weighting response maps produced by the both correlation filters.The experimental results on the OTB2013 dataset demonstrate that the proposed method has improved the tracking accuracy and robustness.(2)A Continuous Convolution Operator Tracker(C-COT)based on deep features is investigated and then improved.To address limited tracking performance resulting from only single resolution feature in the conventional DCFs,C-COT converts feature map of each convolutional layer into a continuous function by implicit interpolation,and defines and learns correlation filters in the continuous domain.It takes the weighted summation of correlation response maps in each layer to get the final confidence map.In this way,the feature maps with different resolutions are merged naturally in C-COT.Based on the C-COT algorithm,this paper introduces a scale adaptive mechanism to handle the object scale variation.Specifically,it performs multiscale sampling on candidate regions for the object,and extracts HOG features to construct scale filters to estimate the object scale accurately.The experimental results show that considering the success rate and precision under different tracking conditions,the proposed method overcomes the scale variation of object effectively and has a significant improvement on tracking performance.