Object Tracking with Shallow Convolution Feature
Wei Wang, Mingquan Shi, Weiguang Li · 2017
Traditional target tracking algorithm uses manual extraction of features, which is difficult to cope with the challenges of rotation and occlusion and deformation. Based on the deep learning method, the convolution neural network is used to extract the features. Because of lost a lot of spatial information in the convolution process, it's easy to make the tracking target drift. In this paper, we use a shallow convolutional network without second training to extract features for tracking, which combines hard negative mining technology and bounding box regression to refine the target location. We have compared our tracker performance with others state-of-the-art tracker. The obtained experimental results in OTB dataset demonstrate the effectiveness of our proposed tracker has outperformed the compared tracking algorithms.