Research On Multi-Model Framework Based Tracking Algorithm

Object tracking aims at enabling the computer to estimate the location and state of the specified target from the videoes like human eyes.The study of object tracking is of great importance in many areas.Considering the change in illumination,occlusion,background clutter,low resolution,low contrast,scale change and other scenes,the use of multi-model can reduce the tracking failures caused by the accumulation of small errors.The tracking algorithm via multiple models using entropy minimization takes use of a tracker and its historical snapshots to constitute a model ensemble,where the best model is selected based on a minimum entropy criterion to prevent model drifting problems.In the traking algorithm via multiple models using entropy minimization,the image patch is processed by a fixed feature mapping method to obtain binary vector with a fixed dimension.However,this method can not distinguish the object and background in a video with similar object and background.So we propose an adaptive binary feature encoding method,the quantization number is calculated based on the dissimilarity between the object region and the surrounding region in the first frame,and the quantization thresholds are decided by the feature clusters of each channel using the K-Means method.This method enhances the distinguished ability between the object and background.The traking algorithm via multiple models using entropy minimization trains SVM classifier with new training data at each frame,however,all models may be wrong updated after a long time.In this paper,we restrictes the model updating conditions with the prior knowledge in first frame and the censorship mechanism.Before using the new training data,we need to compare the updated model with the previous one and the model of first frame to avoid unnecessary model updates.If the updated model is similar with the previous one and is not similar with the first frame one,we skip the update at this frame.The experiments show that our tracker based on multi-model framework,using the adaptive feature encoding method and restricting the model updating conditions,is robust in the case of low resolution,low contrast and occlusion.