The Study Of Object Tracking Algorithm Under Occlusion And Deformation

Object Tracking plays an important role in computer vision and also is a significant component in many practical applications such as robot vision,satellite navigation,unmanned vehicle and human-computer interaction.Although much progress has been made in object tracking field in these years,it is still very difficult to design a good tracking algorithm for all scene which largely because the current algorithm can't deal with the sudden changes of target object very well.There many reason could cause the sudden changes of target object such as occlusion,deformation of target,illumination variation and background clutter.In order to improve the robustness of the tracking algorithm,we must enable the object tracking algorithm to handle the sudden changes of the target very well and at the same time we can't decline the real time performance of the algorithm.If the real time performance of the tracking algorithm is poor,that algorithm can't not be used widely.This paper proposed some methods to improve the robustness and real time performance of the tracking algorithm based on previous work,aiming at the above problems.Our main works are as follows:(1)Methods for occlusion detection during trackingWhen the target suffering occlusion,the appearance of target would have a significant change and in tracking based correlation filter this change would be reflected by the decrease of the response curve.More significant the change then there would have a more dramatic decrease in the response curve.This property can be used to detect the the appearance of the target.When we know the target has suffered some appearance change,we also need to know the specific reason that causes this change.Here we design a method called back-track to detect whether this appearance change is caused by occlusion.We assume that the occlusion gradually close to the target.The extensive experiments shown that the proposed methods are valid.(2)Method for deformation detection by deep learning features in trackingThe deep learning features used here are extracted from CNNs.There are many layers in CNNs and the output of every layer can be seen as a type of feature of the original image.The features extracted from the shallow layer are similar with the original image which contain many spatial details.The features extracted from the deeper layers contain more semantic information which are more insensitive to the deformation of the target,the rotation of the target and illumination variation.Therefore the two types of features can be used to detect the deformation of the target.(3)Method for improving the real-time performance of the target tracking algorithmThe application areas of object tracking are wide such as robot vision,satellite navigation and human-computer interaction.There exists a common requirement in these application areas which is the algorithms used in them must have high real time performance.For example,in human-computer interaction,if the algorithm does not have enough real-time performance,the user experience would suffer.However the real-time of most tracking algorithm is poor,thus it is very meaningful to improve the real-time performance of the tracking algorithm.Aiming at this problem,we proposed a general tracking model then analyse the parallelism of this model and finally we give a parallel optimization scheme for this model.The experiment have shown that the parallel optimization scheme is valid.