Long-time Object Tracking Based on Hierarchical Convolution Features

摘要

Traditional object tracking methods, mainly based on manual features (e.g., histograms of oriented gradients, color names and color histograms), have limited ability to complex scenarios. As convolution neural network has achieved great progress on image classification, it has been proved to have strong feature extraction ability and superiority to traditional object tracking methods. In this paper, we propose an improved object tracking method, where the pre-trained convolution neural network is used for hierarchical feature extraction. Moreover, a robust model updating strategy and an object re-detection strategy are introduced to our method. Several experiments on public data sets are provided to demonstrate that our method indeed improves the performance in illumination changing, occlusion, low resolution, while enhancing the overall accuracy and success rates.