Target tracking method based on online state learning and estimation

Abstract

The invention provides a target tracking method based on online state learning and estimation and belongs to the computer vision, computer graphics and image technical field. According to the method, a target positioning and state estimation network is obtained; and the target positioning and state estimation network is composed of a feature extraction network and a regression network, wherein the feature extraction network is a pre-trained AlexNet network, and the regression network is a recurrent neural network (RNN). In an initial network training process, an initial training set and a stochastic gradient descent method are utilized to train the target positioning and state estimation network, after being trained, the target positioning and state estimation network obtains an initial ability to perform target positioning and state estimation. In a tracking process, the target positioning and state estimation network performs forward processing on an inputted image and directly outputs target related information corresponding to the image, wherein obtained target probability and state information decides whether the network to perform online learning, and target position and size information can realize the localization of a target, and therefore, the tracking of the target can be realized.

Description

technical field [0001] The invention relates to computer vision and the technical fields of computer graphics and images. Background technique [0002] Visual object tracking is an important research topic in the field of computer vision. Its main task is to obtain the continuous position, appearance and motion information of the object, and then provide the basis for further semantic analysis (such as behavior recognition, scene understanding, etc.). Target tracking research is widely used in intelligent monitoring, human-computer interaction, automatic control systems and other fields, and has strong practical value. At present, target tracking methods mainly include classical target tracking methods and deep learning target tracking methods. [0003] The classic target tracking methods are mainly divided into two categories: Generative Methods and Discriminative Methods. The generative method assumes that the target can be expressed through some kind of generative proce...

AI Technical Summary

Problems solved by technology

[0005] In order to be able to adapt to changes in the target, tracking methods usually need to update the corresponding model online. However, the changing state of the target appearance will greatly affect the accuracy of model learning. State-insensitive learning will cause tracking drift due to accumulating learning errors

Images