Research On Object Tracking Based Temporal-Spatial Model

Object tracking is one of the hot topics in the field of computer vision.Because of the importance of object tracking in video surveillance, visual navigation, human-computer interaction and augmented reality, it attracts a large number of scholars who are devoted to the research of object tracking algorithms. Object tracking is affected by many factors, such as illumination change, deformation, occlusion, fast motion, background disturbance, so it is difficult to propose a tracking algorithm to deal with the complex scene. The research of object tracking algorithm still has very important significance.Aiming at the problems of model drift and background interference in object tracking field, this paper proposes an object tracking algorithm based on temporal-spatial model. The drift is corrected using the historical model of the object, and the position of the object is predicted using the motion correlation between the object and the surrounding objects. Finally,a more robust tracking algorithm is obtained by fusing the tracking model in temporal and spatial domain. The main contents and innovations of this paper are as follows:(1) A multiple instance learning tracking algorithm for balancing the weight of positive and negative samples is proposed. The multiple instances learning tracking algorithm selects the best weak classifier from the pool of weak classifier based on maximizing the log likelihood function of the sample packet, and then obtains the strong classifier tracking model through the linear combination of the weak classifiers. However, in practice, positive and negative samples are distributed in different states,leading to the choice of weak classifier, the role of the sample is negligible.Therefore, this paper improves the performance of the multiple instances learning tracking algorithm by balancing the weights of the positive and negative samples.(2) A tracking algorithm for the correction of model drift based on history models is proposed. In the tracking algorithm, in order to adapt to changes in the object, model updates are indispensable. But the model update will inevitably introduce errors into the algorithm, causing model drift. This paper constructs a model pool by preserving the historical model of the object, and then selects the object model most suitable for the current frame from the model pool to correct the model drift.(3) The algorithm of using spatial information to assist object tracking is proposed. The algorithm is prone to drift when the object is in a mixed background, or when it is interfered by similar objects. The spatial correlation information can be used to locate the object. The object and other objects around may have relatively fixed motion correlation, and the position of the object can be predicted by the relative positional relationship between the object and other objects.