EFTrack: A Lightweight Siamese Network for Aerial Object Tracking

Abstract

Visual object tracking is a very important task for unmanned aerial vehicle (UAV). Limited resources of UAV lead to strong demand for efficient and robust trackers. In recent years, deep learning-based trackers, especially, siamese trackers achieve very impressive results. Though siamese trackers can run a relatively fast speed on the high-end GPU, they are becoming heavier and heavier which restricts them to be deployed on UAV platform. In this work, we propose a lightweight aerial tracker based on the siamese network. We use EfficientNet as the backbone, which has less parameters and stronger feature extract ability compared with ResNet-50. After a pixel-wise correlation, a classification branch and a regression branch are applied to predict the front/back score and offset of the target without the predefined anchor. The results show that our tracker works efficiently and achieves impressive performance on UAV tracking datasets. In addition, the real-world test shows that it runs effectively on the Nvidia Jetson NX deployed on DJI UAV.