Moving Object Tracking Based On Discriminative Correlation Filters

Moving object tracking is a fundamental aspect of pattern recognition,which has been an essential component in fields such as biology vision,self-driving vehicle and video surveillance.Many methods and techniques from machine learning and artificial intelligence have been adopted in object tracking that have greatly improved object tracking methods.Despite the many years of development in this area,object tracking remains a challenging and complex problem due to many sources of interference,including appearance variations,occlusions,background clutter,and illumination changes.The theory of discriminative correlation filters was originally developed in signal processing and has been integrated into object tracking to achieve both high accuracy and high speed.As a result,correlation-filters-based tracking approaches have attracted many researchers and have become a hot topic in object tracking research.Although object tracking based on discriminative correlation filters has achieved great improvements compared with traditional tracking methods,many issues of correlation-filters-based trackers,such as model learning,object scale estimation,context awareness and deep learning,have to be further researched and further solved to enhance the overall tracking performance.To address these problems,this work is under the discriminative correlation filters framework and focuses on aspects such as object appearance model learning based on self-paced learning,object scale estimation in a continuous space,learning context-aware discriminative correlation filters for object tracking with adaptive regression targets and end-to-end multi-context Siamese network with residual hierarchical attention for object tracking.The main research contents and contributions of this work are as follows.1)This work researched on self-paced-learning-based appearance model learning method and proposed a novel object tracking algorithm based on self-paced learning.When learning a new appearance model,many existing tracking methods generally ignore the reliability of tracking results,which may integrate corrupt samples into model learning.Although some algorithms have established criteria for selecting samples,such artificially established hard criteria may not meet model's requirements for training samples.In this work,we propose a novel tracking method that integrates the learning paradigm of self-paced learning into object tracking such that samples that are reliable for model learning can be automatically selected.In contrast to many existing model learning strategies in object tracking,we discover the missing link between sample selection and model learning,which are combined into a single objective function in our approach.We formulate a new self-paced function with a mixture scheme that is a hybrid of hard weighting and soft weighting.On the one hand,compared with the soft-weighting scheme,the mixture scheme is fault tolerant.Compared with the hard-weighting scheme,on the other hand,it can assign real-valued weights reflecting the latent reliability of samples in training.Moreover,we analyze the characteristics of object tracking and take them into account by integrating a constraint vector into the self-paced function.We compare our proposed tracker with other state-of-the-art trackers on OTB dataset from the aspects of quantitative comparisons and qualitative comparisons.The experimental results demonstrate the robustness and efficiency of our tracker.2)This work researched on object scale estimation in a continuous space and proposed a novel object tracking algorithm that is able to estimate the object scale in a continuous manner.A mature scale estimation method can greatly improve tracking performance and provide accurate target information for model training.However,some object tracking approaches ignore the scale estimation problem or adopt a heuristic and exhaustive scale-estimation strategy.Such practices greatly limit the precision of scale estimation.In this work,a scale estimation equation is deduced based on the classifier response in each frame.Then,the adaptive target scale can be calculated mathematically rather than heuristically producing some samples with pre-designed criteria.In addition,to consider the prior knowledge of object tracking and obtain a stable scale variable,we design an iterative strategy with a constraint function to update the scale variable and the bandwidth of kernels.Moreover,we formulate a sample learning scheme with dynamic thresholds that enables model training based on the average losses of previous frames.We demonstrate the robustness and efficiency of our tracker on the OTB dataset,and analyze our tracking results from the aspects of quantitative comparisons and qualitative comparisons.3)This work researched on joint context awareness and regression targets adaptation,and proposed context-aware correlation filters for object tracking with regression targets adaption.Context information includes the background information around the target,which can provide useful clues for locating the target among consecutive frames in a video clip.However,most correlation-filters-based trackers are not context-aware or consider only limited contextual information about a target object,mainly because the search area in every frame is a small neighboring area of the target object,and the use of a cosine window to address the boundary effects further deceases the effective contextual information.In this work,we formulate a new objective function based on discriminative correlation filters that exploits the missing link between context awareness and regression targets adaptation.Moreover,we add the contextual samples around the target object to our proposed objective function as a regulation term.Therefore,the contextual information can be taken into consideration and included in the learned correlation filters by obtaining the closed-form solution of the objective function.Furthermore,in contrast to most traditional correlation-filters-based methods that use static regression targets,we can learn the adaptive regression targets jointly from frame to frame.In our algorithm,we propose a constraint matrix that includes the motion and distribution information of the target in the current frame and is constructed to obtain new regression targets;thus,the regression targets are able to continuously adapt to the target.We execute our proposed tracker on the OTB 50,OTB 2013,OTB 2015 and VOT 2017 datasets and compare the results with those of 18 and 7 state-of-the-art trackers to demonstrate the efficiency of our tracker on OTB and VOT,respectively.Different components in the proposed tracker are analyzed through ablation studies.4)This work researched on residual hierarchical attention in Siamese convolutional network and proposed end-to-end multi-context Siamese network with residual hierarchical attention for obj ect tracking.Many methods from deep learning have been adopted in object tracking that have greatly improved deep-learning-based object tracking methods.However,some deep-learning-based trackers utilize pre-trained networks,such as VGG and AlexNet,and then attach other existing tracking approaches.This practice does not implement end-to-end training,and does not fully utilize the power of deep learning.Furthermore,the running speed of many deep-learning-based trackers is too low to meet real-time requirements.In this work,we propose a novel multi-context Siamese convolutional neural network that uses a residual hierarchical attention mechanism to achieve high-performance object tracking.In our end-to-end network,we propose a multi-context correlation filters layer to enhance the discriminatory capability of the proposed tracker,which jointly considers context awareness and regression targets adaptation during object tracking.This network is trained offline in an end-to-end manner and is capable of performing real-time tracking.Moreover,we propose residual hierarchical attention learning that uses residual skip connections in the attention module to produce a more efficient feature.The attention module extracts hierarchical attention maps from different upsample layers;we analyze the different characteristics of these attention maps for object tracking.Then,we sum them to obtain a more generative attention-aware feature.We execute our proposed tracker on the OTB 50,OTB 2013,OTB 2015 and VOT 2017 datasets and compare the results with those of 20 state-of-the-art trackers on OTB and 7 state-of-the-art trackers on VOT to demonstrate the efficiency of our tracker.In ablation studies,we analyze different components in the proposed tracker.In summary,this work has carried out in-depth researches on moving object tracking based on discriminative correlation filters,and obtained good experimental results.This work provides new solutions for the research on object tracking.