Real-time Video Object Tracking Algorithm Research Based On Detection Feedback

With the progress and development of society, science and technology has been entered into daily life, especially the digital visual technology plays a more and more important role in the daily life. The image processing and the object tracking have played a decisive role in the visual technology. In recent years, the object tracking in the image processing has become one of research hotspots for it have a wide application prospect and trend in development in the traffic surveillance, human-computer interaction, community safety and other commercial civil, military fields. The object tracking algorithm based on detection feedback which also called the discriminant object tracking is the mainstream of current tracking algorithm. The traditional generative model doesn’t consider the influence of background which is easily affected by environmental factors. However, discriminant tracking methods can extract distinguishing features in target and background regions. At same time, online training classifiers tracking problem proposed as a classification problem which provides a new direction for the study of object tracking.In this thesis, we research the object tracking from two aspects, the first part of the model use the popular subspace learning method. At most time, the image dimension is large and target itself contains some useless information so the key factors of target tracking is how to choose the distinguish features information. And the subspace learning tracking use the target information and the background information at the same time to construct an over complete dictionary and use it to obtain the sparse representation of foreground and background as the tracking features. It achieved a robust result, finally. The second part we research the practical application of tracking based on detection feedback which expand the usage of tracking object, this part mainly research the hand tracking, because the hand is not rigid and apparent changes. So the key and the main research direction is how to track hand in the complex environment.Our work and new contributions can be concluded as:1. We propose a tracking algorithm based on multi-block and sparse representation to solve the occlusion problems which is easily occurred in the process of object tracking. The algorithm constructs an overall model from two aspects, the first part of the model using the global information to construct discriminant tracking, this part using the foreground and part of the region far away from the target area as the background to build the complete dictionary and use these to obtain the sparse representation of foreground and background which treat as the training data of the original classifier. The classifier we choose the linear SVM, the new frame construct each particle classifier score base on the particle filter framework, the highest score candidates as the most similar to the object tracking region. The second part, we construct the object region and background region base on local information use the multi-block and calculation of reconstruction error and allocation different weights of each block according to reconstruction error. Finally, we design the similarity measure and the whole model update strategy. Through compare with the mainstream tracking method in open tracking test set, the proposed method has better accuracy and robustness.2. We track the hand in the complex environment using the characteristic of the hand. Through using the color information and saliency information to connected the high-level information and the underlying information which can acquisition hand area roughly. It can reduce the searching region for the subsequent hand tracking. Then we using Hough tracking and optical tracking detect hand region online. At same time, the back projection, optical flow and graph cut are used to obtain hand region and achieve the hand tracking.