Research On Object Tracking Based On Deep Learning And Correlation Filters

As one of the most important tasks in the field of computer vision,object tracking has been widely used in robotics,human-computer interaction,autonomous driving and 3D reconstruction.The aim of the object tracking is to estimate the trajectory of a target in a video,given only its initial location.Since the selected target can be any region of interest,this type of tracking is called model-free tracking.Although much progress has been made in this field,it is still a challenging topic to design accurate and efficient object tracking algorithms considering the tracking occlusion,fast motion,background clutter and illumination changes.In recent years,the correlation filters based tracking methods have grasped researchers' attention due to their promising performance and computational efficiency.How to combine deep learning with correlation filters for efficient object tracking is studied in this thesis after analyzing the deep learning based and correlation filters based tracking methods respectively.The main contributions of this thesis are summarized as follows:1)A framework combining the Convolution Neural Network(CNN)and particle filters is proposed for object tracking.In particular,the traditional hand-crafted feature,HOG,in particle filters is replaced with deep feature extracted by CNN to enhance the tracking feature robustness.In addition,to better analyze the impact of different depth features on tracking performance,the features from different layers of AlexNet are compared on some popular tracking benchmarks.Experiments verify the effectiveness of deep feature for object tracking.2)An end-to-end lightweight network is designed to learn deep features and perform the correlation tracking simultaneously.Specifically,Discriminant Correlation Filter(DCF)is added into a Siamese network as a special correlation layer,which enables the learnt deep features tightly couple to the correlation filter.3)A framework is presented to integrate the depth and appearance features into the Correlation Filtering(CF)architecture.This framework can take the advantage of both depth and appearance feature by fusing them on the feature level and feeding the fusion feature to correlation layer.This method can achieve a state-of-the-art performance with a high speed(100 fps)on OTB-2013 and OTB-2015 benchmarks.