| Visual target tracking is an important branch of computer vision.It is widely used in artificial intelligence systems,such as video monitoring system,intelligent traffic system,intelligent robot navigation,human-computer interaction and precision guidance system.With the rapid development of AI,more and more researchers pay attention to the research of target tracking algorithms.Although many excellent tracking algorithms have been proposed,they cannot be fully adapted to the complex evironments such as illumination change,attitude change,fast motion,target occlusion and background clutter.Therefore,in view of complex environment scenarios,we propose robust tracking algorithms to deal with these challenges.The main contents of the work are listed as follows.1.This paper proposes a collaborative model by incorporating the local and holistic models together which is corresponding to discriminative and generative models in tracking-classification-framework.First,at the local level,an on-line Random Forest(RF)classifier is trained to distinguish the superpixels of the object from the background.A series of local superpixels are used to represent the target,so as to adapt the appearance variances.The discriminative model is used to classify superpixels in the next frame as either belonging to the object or background.A confidence map with dependability and stability is formed to measure the probabilities of superpxiels pertaining to the target from the classifier.A modified mean-shift is proposed to work on the confidence map to find the peak,where is the position of the target.Meanwhile,a separate component for managing the training set dynamically is employed to control the updating of the RF model.Then,at the global level,the target is expressed by extracting the covariance matrix of the target with multiple scales and matching the target global covariance feature of the last frame to determine the target scale.Finally,the experimental results show that the proposed algorithm is effective and robust.2.This paper proposes an improved Compressive Tracking(CT)algorithm based on Local Sensitive Histogram(LSH).It significantly improves the conventional CT in four aspects.First,the effcient illumination invariant features extracted on the basis of the LSH are used to represent the appearance of a target,which is robust to illumination changes.Second,the color attributes tracker is adopted to predict the target position for re-building the new weighted discriminant function which brings the color information to make up for the inadequacy of Haar-like characteristics.Third,a new model updating mechanism is proposed to preserve the stable features while avoiding the noisy appearance variations during tracking.Fourth,a trajectory rectication method is employed to refine the tracking location when possible inaccurate tracking occurs.Finally,experimental results conducted on benchmark dataset show that the algorithm can solve the difficult problems of illumination change,target occlusion and appearance change in complex background.3.This paper proposes a collaborative model based on complex cells and keypoints and focuses on three key factors: an effective representation to consider appearance variations,an effective application of the keypoints and an incorporation of contextual information.First,due to the key point is the ideal local expression,the key points and optical flow are used to track the target roughly.Second,since complex cells can effectively explore multi-scale contextual information,complex cells are further fused to locate targets accurately.Then,the internal cell matching is used to measure the degree of appearance change.The adaptive learning rate parameter is used to update the appearance model of the target to avoid noise interference.Finally,the experimental results show that the algorithm can well solve the influence of occlusion,scale change,fast motion and target deformation.In conjunction with tracking acceleration modules,the proposed algorithm performs in real-time. |
PDF Full Text Request