Research On Object Tracking Algorithm With Superixel And Compact Feature

Object tracking in the video is one of the important contents in the field of computer vision research,and has attracted much attention.Due to the problems such as occlusion,scale change,shape deformation,fast motion,complex background and low resolution,the existing tracking algorithm can not solve all the problems and there are still many defects.Thus it is this paper’s purpose to resolve those tracking problems when the object is in the e-vent of occlusion or scale changes,through improving the existing object tracking algorithm.The main work are summarized as follows:In order to solve the problem of superpixels clustering error in the target and background region,a new algorithm based on multiple nearest neighbors is proposed.According to the difference of pixels numbers inside and outside the selection frame,the algorithm determines the symbol of target-background confidence value of the superpixel.The target-background confidence value of the superpixel feature keep unchanged or reduced,is determined by whether the symbols of the target-background confidence value of the superpixel and the cluster that belongs to the superpixel feature are the same or not.In the view of the distance between the multiple nearest neighbors clustered and the feature,we determine the weight of the neighbors which are getting involved in calculation of the weighted sum.The sample with the highest confidence value is our target.Based on the cumulative error of tracking the training samples,the best number of nearest neighbors is determined.Aiming at dealing with the problem of non-target appearing in the selection box,a superpixel tracking algorithm based on projection compression and target recutting is proposed.During the phase of training or appearance model updating,the target region or best state with the watershed algorithm is sliced,the superpixel features with higher confidence value of the selection box labeled as target and others as background are added into the feature pool.The features are clustered and then the confidence value of cluster and the elements is calculated.The confidence value of the elements clustered correctly is remained unchanged,and the absolute confidence value of the elements clustered wrongly is reduced.The random forest and the number of samples weak classifiers are trained with three sample sets.The first set is selected from the area in the inscribed circle.The second set is chosen from the area between the inscribed circle and the circumscribed circle.The third set is got from the area between the circumscribed circle and the circle with a radius that is 2 times the circumscribed circle’s.The three sets are subjected to multi-scale filtering and then convoluted to obtain the feature of the sample block after projection compression,and the classifiers in the case of different occlusions.During tracking phase,the center of the former frame is set as the current center,within the double area of the tracking box for Gaussian sampling.According to the clusters that the best number neighbors belongs to,the cluster is determined and the confidence value is calculated.Then the sample’s confidence value is calculated,the sample with the biggest confidence value is selected as a tentative target.According to the tentative target and the occlusion index,the current occlusion situation is determined,and the corresponding classifier is selected.The maximum likelihood estimation of the conditional probability corresponding to the candidate sample is calculated.According to the relationship of weight between relevant confidence value and the conditional probability maximum likelihood estimation,the final candidate sample is determined,then target tracking is achieved.Experiments show that in the case of occlusion and scale changes,non-target appears in the selection box,the new method can track the object well.