Research On Data Imbalance In Visual Tracking

Visual object tracking is one of the most important and basic tasks in the field of pattern recognition and computer vision.Not only can it cooperate with other algorithms to complete certain tasks,but it can also be used as the basis for many advanced computer vision tasks.In generic visual tracking,the loca-tion of the target is given in the first frame,and the tracking model estimates the locations of the target in the following frames in sequence.Therefore,the biggest challenge for the tracking task is how to use a limited number of training samples(only from the first frame of the video)to obtain a tracking model with generalization capabilities.Most tracking algorithms treat visual tracking as a binary classification task,in which the target given in the first frame is defined as positive class and the background is defined as negative class.To improve the discriminative ability,existing tracking algorithms focus on enhancing the feature representation and designing the mechanism of model updating.How-ever,limited attention is paid to the data imbalance issue in the visual tracking task.To comprehensively study the impact of data imbalance in the visual track-ing task and find effective solutions,this thesis conducts the research along two lines:one is related to the key components of tracking algorithms and another is related to the two types of data imbalance issue.We focus on the specialty of visual tracking when studying the data imbalance,and optimize the tracking al-gorithms from the perspective of solving data imbalance.We carry out relevant research,and achieve the following innovative results:(1)For the imbalance of positive and negative samples,we propose an en-hanced model initialization method based on progressive learning.Based on the experimental observations,we heuristically propose to decouple the nega-tive samples into four fine-grained types.Afterwards,according to the order of negative samples from easy to hard,the positive samples and each type of nega-tive samples are exploited to train the model in a multi-stage manner.Through the training method above,the model can stably obtain the ability to distin-guish between positive samples and negative samples of a certain type in each training stage.Due to the progressive learning method,the discriminative abil-ity of each stage of the model is improved on the basis of the previous stage,which guarantees the training efficiency of the model and the final discrimina-tive ability.The visualization of the feature distribution proves the effectiveness of the proposed algorithm,and the model achieves excellent performance on the benchmark datasets.(2)For the imbalance of easy and hard samples in visual tracking,we pro-pose the weighted-gradient loss,which calculates the weighting coefficients of corresponding samples using the gradients of samples in the back-propagation process of model training.It can help to overcome the difficulty of defining easy and hard samples manually.Also,it balances the contributions of easy and hard samples to model optimization.It prevents the easy samples from dominating the optimization process of the model,and improves the discriminative ability of trackers under difficult conditions.(3)A novel modeling method for visual tracking is proposed-the track-ing algorithm based on multi class classification.In the proposed method,five classes of samples are defined according to context relationship between the target and the background,instead of defining two classes of samples accord-ing to the overlap ratios between the target and samples as in existing tracking algorithms.Experiments shows that the multi class classification model helps improve the tracker's perception of the context between the target and the back-ground,and improves the discriminative ability of the model and the accuracy of target scale estimation.(4)An adaptive candidate generation strategy is proposed.Compared to the random candidate generation strategy in existing tracking methods,the pro-posed strategy exploits the relationship of the target and the background con-tained in outputs of multi class classification model to guide the search direc-tion of the target.The proposed method can not only generate more target can-didates of high quality,but also get more target samples for online updating to alleviate the impact of data imbalance.The performance of the proposed methods above is evaluated on benchmark datasets and compared with state-of-the-art trackers.The experimental results prove the effectiveness of the methods proposed in the thesis.