Multi-domain learning target tracking algorithm based on objective regression optimization

Abstract

Convolutional Neural Network (CNN) is widely used in target tracking for the computer vision, where Intersection of union (IOU) is the most popular evaluation metric in the target detection criteria, but IOU cannot be optimized for tracking algorithms in the case of non-overlapping bounding boxes. GIOU can be optimized for tracking in the case of non-overlapping bounding boxes, but the slow convergence speed of GIOU leads to inaccurate detection, which results in low tracking accuracy. To solve the above problems, a DIOU-based MDNet tracking method is proposed in this paper. In order to solve DIOU loss does not have a penalty term for the aspect ratio of the target box, we propose CIOU-based MDNet and experiments show that the accuracy of this method is improved by 3% compared with MDNet trained with traditional IOU, GIOU or DIOU.