Improved MDNET Tracker in Better Localization Accuracy

Abstract

Object tracking is one of the most important issues in the field of computer vision, which has many applications in the automotive, defense, robotics, medicine, industries, etc. In recent years, various researches have been done in this field and due to its many applications, research in this field continues. In this paper, a deep end-to-end, multi-domain method, or MDNET was examined. The MDNET method works well in tracking video sequences, but has the following problems: The first problem is finding the target position, where the candidate is selected as the target with the highest score. To solve this problem, a new way to find the target position is provided. In this way, the average of the top five candidates is selected as the target position, which was able to increase the IOU standard of the basic method on the OTB100 dataset from 69 to 70.3 and also the results based on the OTB50 and the various challenges are investigated. The second problem is the loss of target in some frames, In this case, instead of drawing the candidates around the target, the candidates are drawn in the whole image and the candidate with the highest score is selected as the target. This method was able to increase the IOU standard on the OTB100 dataset from 70.3 to 72.2 and the results based on the OTB50 and the various challenges are investigated.