| With the rise of artificial intelligence technology,computer vision,natural language processing and speech recognition technology have been brought into thousands of households.The visual single-object tracking task emerges in the field of computer vision,in order to obtain object's real-time location and to stably track the object for the long term.The visual single-object tracking is one of the basic tasks in computer vision technology,which can be applied to researches such as video analysis,human-computer interaction and automatic driving.Therefore,the visual single-object tracking has been considered as an important issue by many visual researchers.In recent years,more and more researchers have used tracking-by-detection to carry out the visual single-object tracking task.Compared with those using the traditional objecttracking algorithm,the tracker utilizing the tracking-by-detection algorithm is faster in tracking and more efficient in data high-precision processing.Its algorithm is also simpler than the traditional one.Some problems such as deformation,occlusion and motion blur may easily occur while performing object tracking.This is due to the complexity and diversity of the environment that the object in the video is located in,and the changes which may happen on the object itself.Meanwhile,how to apply the object-tracking algorithm into the real-scene is also a question worth exploring.Thus,in response to the above issues,this paper mainly involves the following innovations:In order to overcome the incompatibility of the motion tracking object when the object tracking is performed by using the tracking-by-detection algorithm based on the principle of correlation filtering,this paper proposes an object tracking algorithm that uses Lasso constraints and integrates optical flow information.The first stage is feature extraction.Optical flow information are incorporated into the channel block when algorithm extracting feature.Then multi-feature fusion is performed after channel block fusion.Secondly,Lasso is used to constrain the object function of DCF tracker.Considering the discontinuity of theconstrained objective function on the domain and the optimizing efficiency of the object tracking,a block coordinate descent algorithm is used to optimize the constrained objective function.The experimental results show that compared with DCF-based visual tracking algorithm,the proposed algorithm can effectively deal with motion blurred objects and achieve robust visual object tracking in motion blurred scenes.Aiming at the object deformation problem,a visual object tracking algorithm based on deep learning Heat-Map model is proposed.Firstly,the influence of the object deformation on the benchmark algorithm when tracking the object is analyzed.It is difficult for the discriminator to find out the position of the current frame object based on the information of the previous frame during the period of fast object deformation.To solve this problem,using deep learning Heat-Map model to re-sense the position of the object after its deformation is proposed.Also,in order to obtain an accurate deformation object position,this paper designs a search area-generation algorithm.It can be utilized to estimate the search coordinate area where the object with different deformation degree in the next frame may be.Then,the generated search area is inputted into the saimese network discriminator,and the coordinates of the object are returned.Also,a responsiveness evaluation algorithm is designed to evaluate the current response graph status.Finally,update mechanism by combining the dynamic Heat-Map information.The experimental results show that the proposed visual object tracking algorithm based on the deep learning Heat-Map model is improved compared with the benchmark algorithm,which is especially suitable for visual object tracking in deformation scenarios.In order to explore the application of object tracking algorithm in real-application,this paper proposes a security application combined with object tracking algorithm.Firstly,the application interface is designed and built.Four modules,which are an interface logic control module,a face-detection module,a face-recognition module and face-tracking module,are designed in the application.The effectiveness and feasibility of the system are verified by experiments.In conclusion,in view of that the visual object-tracking task is significantly important in researches for video analysis,human-computer interaction and automatic driving,this paper studies in two key object tracking problems.What is more,this paper utilizes theoptimization theory and the properties of deep Heat-Map features to solve the above two key problems.Two object tracking algorithms are proposed.Experiments are carried out on the real scene and on the public evaluation algorithm data set.Moreover,the effectiveness of the proposed algorithm is verified as well.This paper provides a new idea for visual object-tracking tasks in solving the issue of motion blur and deformation.Besides,the application of object tracking in actual intelligent video surveillance is also explored. |
