Research On Face Tracking Based On Particle Filter

Face tracking is an important branch of the video interactive technology.It has avery widely applications in human-computer interaction,face recognition,robot visionand video surveillance.Face tracking which is detecting whether there is a face in theimage sequence,if it exist,extracting and tracking it. The focus of face tracking is howto accurate detection of faces,accurate and stable tracking it in a video.In the face detection, In order to solve the low accuracy of the detection speed inthe traditional Adaboost algorithm, this paper presents an modified Adaboostalgorithm based on skin color to improve the accuracy of the face detection. First, thealgorithm rules the out the most background area quickly,reduce the amount ofcomputation and improve the detection speed by using the skin features. And theHaar feature of the face is pick up by the using the rapid integration figure in the skincolor of area. Then the threshold setting method improves the traditional Adaboostalgorithm.The optimal classifiers from every detection are cascaded the ultimatestrong classifier. According to the Haar characters distributions from the strongclassifier, the face area is detected. The experimental results show that the algorithmis effective and the accuracy of the face detection is improved.This algorithm can beeffectively used both in single face detection and more face detection.In order to solve the problem of lowly stability by using traditional particle filterin tracking.This paper used kernel function in the step to calculate the histogram.Thismethod Effectively imorove stability of face tracking by ways of using weighted colorhistogram.In order to solve the problem of low robustness in the tracking based on skinon the influence of illumination intensity changes. In this paper Used contour as thetracking clues.Using contour tracking for a frame when the color tracking isunstable,then back to use the color tracking. Experiments show that this method canstably tracking the face in the case of face rotation, illumination intensity changes andPartial occlusion.The proposal distribution of traditional particle filter directly instead of prioriprobability density,it has a lager gap with posteriori probability desity. In some highprecision occasions，it is easy to lose the target.So this paper introduces a improvedGM(1,1) model. Combined with the predicted results to generate proposal distributionto make the proposed distribution is more close to the posterior probabilitydensity.The experiments shows that this algorithm can achieve effective tracking inthe case of face is complete blocked for a few fames.Finally, This paper according to the weight of particle to sampling in the step ofimportance sampling.This method sampling more particles from the particle withlarger weight and less from it with lower weight.This measures effective solution tothe problem of the degradation of the particles in the tracking process.