| Visual object tracking is the focus of computer vision in the field of artificial intelligence,which is widely used in various fields such as military defense,transportation,medical diagnosis and civil security etc.It works with the goal of predicting the location and scale of the target in the subsequent frames of a video sequence,only initializing the target state in the first frame.It is still a challenging problem to design a precise and robust visual tracking algorithm due to the lack of training samples and appearance changes of the target caused by factors such as occlusion,fast motion,drastic deformation,rotation and background clutters and so on in the complex reality tracking scenario.Since 2013,the correlation filter tracking algorithm proposed by using the correlation concept in signal processing theory has successfully attracted the attention of the majority of researchers.This method transfers complex convolution operation in time domain into sample element-wise multiplication in Fourier domain using Fourier transform and obtains high tracking accuracy while maintaining extremely fast speed,which becomes the hot research direction of visual tracking in recent years.Although excellent tracking results have been achieved in several benchmarks and tracking performances have remarkable improvements compared to traditional tracking algorithms,the correlation filter tracking algorithm still has poor performance due to accuracy reducing and anti-interference ability weakening of its appearance model in complex tracking scenarios.For this reason,the multi-cue fusion strategy is introduced into the correlation filter tracking framework and we carry out in-depth research from the multi-cue fusion perspectives of feature-level and decision-level,which apply complementary cues(including: convolutional neural network features,color histograms,and multi-scale filter templates)to enhance the accuracy and anti-interference of the appearance model and improve the ability of distinguishing target from background.The main research works and innovations of this thesis are as follows:(1)According to the feature-level multi-cue fusion strategy,by combining correlation filter with CNN features,a correlation filter-based tracking algorithm via adaptive weighted CNN features fusion is proposed.Since hand-crafted features(such as HOG and CN)are incapable of capturing the essential invariance information of targets,it suppresses the performance of traditional correlation filter algorithms.Hierarchical correlation filters with lower,middle and higher layers of CNN features are constructed respectively by using the rich details and semantics of CNN features.In addition,adaptive weighted fusion of the correlation response maps is carried out to improve the accuracy and robustness by taking the trackability and motion consistency of CNN features as constraints.To solve the issue of scale estimation in traditional correlation filters,one dimension scale correlation filter is added to our tracking framework.On 50 challenging sequences,compared with KCF,the distance precision rate of our tracker is improved by 19.4%,and the overlap success rate is improved by 21.0%.(2)According to the feature-level multi-cue fusion strategy,by combining correlation filter with color histogram,a structural patch response map fusion tracking algorithm based on correlation filter and color histogram is proposed.When the target suffers partial occlusion or drastic deformation,the global tracking model is unable to successful track the target as it only describes the holistic information of the target,while local tracking model which focuses on the local information of the target can provide reliable tracking cues.Besides,single target appearance model and simple linear combination of local structural responses are insufficient in dealing with a variety of complex challenging scenarios.To this end,two local component trackers: the correlation filter-based structural patch tracker and the color histogram-based structural patch tracker are constructed,which aims at adaptively merging their local responses respectively according to the corresponding loss measurement.Then based on the response confidences of two component trackers,we selectively combine their response maps to generate the final response map.In addition,once the tracking fails,an online SVM re-detection mechanism will be executed to restore the accurate location of the target.On OTB2015,compared with Staple,the distance precision rate of our proposed tracker is improved by 5.2%,and its overlap success rate is improved by 5.0%.(3)According to the decision-level multi-cue fusion strategy,by combining multi-scale correlation filter models with different spatial sizes,a cascaded fusion tracking algorithm based on correlation filter is proposed.Experimentally,correlation filters with different spatial sizes will have great differences in tracking performance when facing with different challenging attributes.And we take advantage of the cascaded framework coming from object detection task to refine the detecting results level by level.A set of correlation filters with different spatial sizes are hierarchically connected in a cascaded mode,and the tracking results of each level are adaptively fused to generate the final result according to their confidence scores.On OTB2015,compared with KCF,the distance precision rate of our proposed tracker is improved by 11.0%,and the overlap success rate of our proposed tracker is improved by 19.6%.(4)According to the decision-level multi-cue fusion strategy,by combining correlation filter with color model,a cascaded-parallel tracking algorithm via collaborative color and correlation filter models is proposed.Since single tracking model has poor versatility,a simple weighted method with single tracking model cannot satisfy various challenging attributes.We thus construct a framework that consists of a cascaded correlation filter tracking component and a parallel color model tracking component to make up for the mutual defects and improve the tracking stability.In the cascaded tracking component,two improved correlation filter algorithms with different scale templates are proposed to coarse-to-fine locate the target step by step.Finally,through evaluating the robustness score of each tracker in cascaded-parallel components carefully,the most reliable tracker is selected for tracking.On OTB2015,compared with MCCT-H,the distance precision rate of our proposed tracker is improved by 0.7%,and the overlap success rate of our proposed tracker is improved by 1.9%. |
