Visual Tracking Based On Multiple Instance Learning

Recently, visual tracking is an active branch in the field of computer vision. It combinescomputer vision, pattern recognition, artificial intelligence and other technical disciplines, andhas broad application prospects. Although numerous approaches have been proposed, robustvisual tracking remains a huge challenge owing to some appearance variations such as objectpartial occlusion, deformation or changes in illumination condition and changing appearancepatterns of the scene. Multiple instance learning performs very well in object tracking. Thispaper proposes two novel methods based on multiple instances learning as follows:Firstly, visual tracking based on particle filter and weighted Multiple Instance Learning(PFWMIL). Particle filter uses Bayesian sequential importance sampling method to estimatethe posterior probability distribution of state variables characterizing a dynamic system.Multiple Instance Learning (MIL) method is recently introduced into the tracking task, whichcan alleviate the target drift to some extent. However, the MIL tracker cannot discriminativelyconsider the importance of sample in its learning procedure. In the base of MIL tracker, weintegrate the sample importance into an efficient online learning procedure when the classifieris being trained, then introduce it in particle filter visual tracking framework. Particle filtermethod is often used to find the location with the maximum classifier score at the new frameand then make the location as the new object location. The proposed method achieves a stabletarget tracking and it solves the drift problems caused by the object and environmentalchanges.Secondly, visual tracking based on compressive sensing and online multiple instancelearning (COMIL). The adaptive appearance models is constructed by extracting sample data，but there does not exist sufficient amount of data for online algorithms to learn at the outset.In order to obtain sufficient data for online learning adaptive appearance model, in the base ofcompressive sensing theories, we extract features by non-adaptive random projections of themulti-scale image feature space. We employ non-adaptive random projections to construct ourappearance model. The tracking task is formulated as a binary classification via a MILclassifier with online update in the compressed domain. The proposed method combines theadvantages of CS and MIL, improves the accuracy of target tracking and solves the driftproblems caused by insufficient sample data.