Research On Robust Object Tracking Algorithm In Complex Scenarios

Object tracking is one of the most important research topics in the field of computer vision,which is widely used in intelligent monitoring,intelligent transportation,human-computer interaction,intelligent medical diagnosis and other aspects.Given the initial state of the target object(such as position,size and other information),the goal of object tracking is to estimate the state of the target in the consecutive image sequence,so as to capture the complete trajectory of the target.Object tracking has made great progress in the past decades.Various excellent algorithms and effective new theories have emerged one after another.However,due to the interference of internal factors(e.g.pose change,scale variation,shape deformation,etc.)and external factors(e.g.illumination change,occlusion,motion blur,background clutter,etc.),it is still a great challenge to achieve robust and accurate object tracking in practical applications.Based on the detailed analysis of the working mechanism of object tracking,in this paper,we carry out relevant research work from the perspective of designing a robust appearance model,and achieve the following innovative results.(1)An object tracking algorithm based on dual-scale weighted structural local sparse appearance model is proposed.Due to the robustness against partial occlusion,pose change and deformation,the local appearance model is favored in the field of object tracking.However,there is a common problem in the existing tracking algorithms based on local appearance model,that is,the different importance of local patches in describing the appearance information of the target is often neglected when the appearance model is established.In this paper,we propose a simple and effective weighted structural local sparse appearance model,which can better describe the appearance information of the target on the basis of taking into account the importance of local patches.Specifically,in this paper,the structural reconstruction errors of local patches are utilized to calculate the patch-based weights that measure the importance of local patches.In the process of constructing the appearance model,the sparse coding of local patches is weighted by the patch-based weights.In this way,the importance of local patches is added to the local appearance model.In order to further improve its robustness,this weighted structural local sparse appearance model is extended to a dual-scale weighted structural local sparse appearance model by multi-scaling.The dual-scale weighted structural local sparse appearance model combines spatial and structural information of different scales,which effectively improves the tracking performance.A large number of experiments on the tracking benchmark dataset show that the proposed algorithm has strong robustness and high accuracy.(2)An object tracking algorithm based on weighted local subspace reconstruction error is proposed.Object tracking is a challenge task as it entails learning an effective model to deal with the changes of target appearance caused by factors such as pose variation,illumination change,occlusion and motion blur.Existing object tracking algorithms adopt either local or global appearance models.The global appearance model can describe the holistic attributes of the target,but it usually regards the target as a whole and treats the whole in the same way.In other words,all parts of the target will be treated equally whether or not the target is occluded or deformed.The local appearance model can capture the local appearance change of the target and deal with different parts differently.However,the local appearance model often lacks the holistic information of the target.The global appearance model and the local appearance model have their own advantages and disadvantages in object tracking,which can complement each other to some extent.Based on this,the global appearance model based on subspace representation and the local appearance model based on local sparse representation are skillfully combined to build a collaborative decision-making appearance model.Specifically,this paper uses the structural reconstruction errors of each local patch after the local sparse representation of the candidate target to design a set of weights,and takes them as penalty factors to adjust each local patch of the reconstruction error of the candidate target under the subspace representation,so as to construct the final observation likelihood.In addition,a model update scheme of occlusion-aware is proposed.The effectiveness of the proposed algorithm is verified by comparing it with some excellent algorithms quantitatively and qualitatively.(3)An object tracking algorithm based on multiple features and fast scale adaptive kernelized correlation filter is proposed.Recently,correlation filter-based discriminative model tracking algorithms have attracted much attention for their high efficiency and strong robustness.However,most correlation filter-based discriminative model tracking algorithms only use a single feature for tracking.Single feature has its limitations in dealing with diverse variations of complex scenarios.At the same time,it is still a challenging problem to realize fast and accurate scale estimation.Finally,correlation filter-based discriminative model tracking algorithms usually update the learned filters with a fixed learning rate to adapt to the latest appearance changes in image sequences when updating the appearance model A fixed learning rate based appearance model update mechanism is suitable for tracking in simple environment or short-term tracking.This paper presents a discriminative model tracking algorithm based on multi-feature and fast scale adaptive kernelized correlation filter.An independent scale filter is introduced to deal with the scale change of the target,and two complementary features are integrated together to further enhance the overall tracking performance.Finally,a dynamic learning rate based model update mechanism is inserted to effectively alleviate model degradation problem by suppressing the influence of noisy appearance changes.Plenty of experiments have been conducted on two large tracking benchmark datasets.Quantitative and qualitative results show that the proposed tracker achieves promising results in terms of tracking efficiency and robustness as compared with other popular trackers.(4)An object tracking algorithm based on global sparse coding and local convolutional features is proposed.Object tracking is a challenging task in many computer vision applications due to occlusion,scale variation,background clutter,and so on.To achieve robust object tracking in complex environment,an object tracking algorithm based on joint appearance model is presented in this paper.The joint appearance model consists of two complementary parts:the discriminative model based on global sparse coding and the generative model based on local convolutional features.In the discriminative model,an effective confidence calculation method is proposed.This method gives the target object a large confidence and other image regions a small confidence,which can well distinct the target from a complicated background.In the generative model,a set of filters are employed to convolve the target region at each position to extract the local appearance representation of the target.Experiments on a tracking benchmark dataset with 50 challenging videos demonstrate the robustness and effectiveness of the proposed algorithm,outperforming many excellent models.