AirFormer: Learning-Based Object Detection for Mars Helicopter
Yifan Qi, Xueming Xiao, Meibao Yao, Yonggang Xiong, Lei Zhang, Hutao Cui
Abstract
In future multiagent Mars detection schemes, the Mars helicopter can assist the scientific missions of Mars rovers by providing navigation information and scientific objects. However, Mars surface exhibits a complex topography with diverse objects and similar textures to the background, posing a great challenge for existing CNN-based object detection networks. In this article, we propose a novel deep learning-based object detection framework, AirFormer, for Mars helicopter. AirFormer embeds a new feature-fusion attention module, MAT, which injects various receptive field sizes into labels. This fusion module is capable of capturing the interrelations between objects with each other while simultaneously reducing computational complexity. In addition, we published a synthetic dataset from the viewpoint of the Mars helicopter: SynMars-Air, which refers to the data collected by the ZhuRong rover. Extensive experiments are conducted to validate the performance of AirFormer compared to SOTA methods. The results show that our method achieved the highest accuracy both on synthetic and real Mars landscapes.