Research Of The Techniques Of Object Tracking Based On Machine Learning

Object tracking is always an research focus wihin the field of computer vision and image processing. It is of great worthy in practice of fields like military spying, security surveillance and smart controlling. It is one of the key techniques of some equipments like weapons and surveillance. In the past decades, a lot of researchers both here and abroad did much research about object tracking. But for the changable of the object appearance, the movable of the object, the complex of the background and the occluding by things near object, object tracking is still a very challenging problem. In recent years, object tracking and pattern recognition using machine learning has been a research focus. Different from traditional object matching, tracking using machine learning theory treat the tracking problem as target classification. Namely, classify the object and background using some algorithm. The object which has the highest probability is chosen as the tracked object, and its position is chosen as the new position of the object. One of a good feature of machine learning is learning, it means that a machine can learn like a man. It can learn the disturbing of the object, like position changing, pose changing and similar objects disturbing, then modulate the tracker to tackle complex tracking problems.In this dissertation, three aspects of algorithm base on machine learning is deeply learned. The main innovation and research results are as follows:In order to solve the problem that the traditional correlation tracking algorithm dose not have good real-time capability and can not adapt object changing, a tracking algorithm using adaptive correlation filter is proposed. The object is tracked in a correlation filtering method. The filter template is trained on-line, then it will be adaptive to the object changing. Weighted the image patch with a cosine window, so that the object near the centre of the image weight more then the object far from the centre. And the weighting will ensure the continuance of the edge between circulant image. According to the circulant marix, convol the image patch and the filter template in Fourier domain. The Fourier domain response is got. The spatial response is calculated through inverse Fourier transformation. The position in the response matrix which has the largest value is the new object position, thus the object is tracked. Experiment on variant videos shows that this algorithm has good matching precision capability and real-time capability. It has good adaptive capability on object shape changing and similar object disturbing.MIL(Multiple instance learning) algrithm is learned. In order to improve the accurate rate and real-time capbility of MIL, a weighted multiple instance learnig algrithm is proposed. First, sample the positive(object) samples and negtive(background) samples according to the object position. Weighted the positive samples according to their similarity to the object. Train the weak classifier set using the feature of the positive and negtive samples. Then, select K best weak classifiers through calculating the maximun value of the log like-hood function of the samples. Weight the selected weak classifiers according their classifying capability. Generate a strong classifier using the weighted weak classifiers. At last, classify the samples sampled from the new coming frame using the strong generated classifier. The classifying result is a posibility. The sample which has the largest posibility value is thought to be the tracked object. Experiment on some videos shows the algrithm can track the object well, and will not be disturbed by similar objects. The new proposed algrithm has better real-time capability than the traditional MIL.In order to improve the real-time capability and accurate rate, combining the compress sensing theory, a tracking algrithm using distance metric learning is proposed. First, sample the positive(object) samples and negtive(background) samples according to the selected object. Compress the Haar-like feature using random projection theory. Then, train the distance metric using the compressed low dimension feature. Last, calculate the Mahalanobis distance between object and the samples sampled from the new coming frame using the trained metric matrix. The sample which has the smallest distance from the object is the object in the new coming frame. Experiment on some videos shows that for the feature dimension is 3/4 less than the origin haar-like feature, the real-time capability improved a lot. Calculating the object position using the trained metric matrix, which means the tracker can adapt its parameters according the object appearance changing, can improve the accurate rate of the tracker.