Abstract:

In recent years, with the development of training data and hardware, a large number of tracking algorithms based on deep learning have been proposed. Compared with the traditional tracking algorithm, tracking algorithms based on deep learning have a great developing potential. However, the traditional convolutional neural network structure cannot effectively perform its powerful feature learning and representation abilities in a tracking task. In this paper, an improved feature extraction network is proposed for video target tracking. Based on the traditional feature extraction network, an attention mechanism and a feature fusion strategy in the form of residual network are introduced. At the same time, a loss function based on the regional overlap rate is introduced in the training stage of the network model, which makes the algorithm produce a better positioning effect. Experimental results show that the improved algorithm can track the target accurately for a long time. Besides, the method has a generalization ability, which can be used for reference for other tracking algorithms based on deep learning.

Key words: attention mechanism, convolutional neural network, residual network, object tracking