Robust Siamese Object Tracking for Unmanned Aerial Manipulator

Abstract

This work introduces Siamese network into unmanned aerial manipulator (UAM) tracking, which is a model-free method and can be applied to track arbitrary objects. Considering severe object scale variation faced in practical UAM tracking application scenarios, a scale-aware Siamese network, i.e., SiamSA, is designed. Specifically, SiamSA employs a novel scale-channel attention strategy to excavate meaningful scale information of the object. Based on this strategy, scale attention network and scale-aware anchor proposal network are constructed to further improve the tracker’s robustness against scale variation. Besides, a new UAM tracking benchmark, namely UAMT100, is developed with 11 attributes and ~35,000 frames on a flying UAM platform. Exhaustive experiments on two authoritative aerial tracking benchmarks and UAMT100 benchmark validate the practicality and effectiveness of SiamSA with a real-time speed. Both the code and UAMT100 benchmark are now available.

Publication
Submitted to the proceedings of the IEEE International Conference on Robotics and Automation, 2022
Guangze Zheng
Guangze Zheng
Research assistant in Vision4Robotics

My research interests include deep learning, visual object tracking, and robotics.