Study On Adaptive Visual Tracking Method

Visual tracking is an important branch of computer vision research area and various kinds of visual tracking methods have been proposed during the recent decades.These methods differ in several aspects such as target object representation, image feature selection,modeling of target object's motion,appearance and shape.In real applications,the performance of visual tracking will be commonly affected by the appearance variation of target objects.Most existing tracking algorithms construct a representation of a target object prior to the tracking task starts,and utilize invariant features to handle appearance variation of the target caused by lighting,pose,and view angle change.It has been proved that invariant feature will not reflect the appearance variation caused by some intrinsic and extrinsic factors.In order to adapt the appearance variation of target objects during the tracking process,the target object is represented by its appearance feature subspace,which will be updated constantly and efficiently during the tracking process using an incremental principal component analysis method.Moreover,in contrast to some gradient descent based method,the posterior probability distribution is propagated by generating a series of samples according to the tracking result of the previous frame,using a kind of particle filter method which derives from Condensation method.Then the tracking result is calculated with an efficient similarity measurement.Experiments demonstrate the effectiveness of our method in the environments where the target object undergoes appearance variation due to object translation,rotation and illumination changes.We notice that in some extreme situations during the experiments,the tacking error rate increases with the existence of particle depletion.A local dual closed loop structure is then introduced to adjust the tracking result obtained from the previous tracking method.It turns out from the experiments that the proposed local dual closed loop structure can enhance the tracking performance not only with the existence of particle depletion but also in common situations.