Scholarlink

As space activities become more frequent, the safety monitoring of Resident Space Objects (RSOs) in geosynchronous orbit (GEO) is of paramount importance. This paper introduces the MT-DeepLabv3+  methodology, which incorporates spatiotemporal characteristics, aimed at enhancing the detection accuracy of RSOs in GEO. The MT-DeepLabv3+ initiates with a single-frame preliminary check of sequential images via a pretrained DeepLabv3+ network, processing multi-scale information through an encoder-decoder structure while considering edge detail information. To address false positives and missed detections in preliminary results, a novel trajectory post-processing technique is proposed. This method effectively identifies genuine target trajectories, eliminating false positives and compensating for missed detections, significantly improving the detection capability for dim and small targets. Verified on the SpotGEO dataset, our research method demonstrated outstanding efficacy, significantly outperforming current advanced schemes with an accuracy rate of 99.52%, reflected in both F1 score and MSE value. Our approach offers promising application prospects in the detection of dim and small targets at long distances and under low-resolution conditions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MT-DeepLabv3+: An Enhanced Deep Learning Approach for the Detection of Resident Space Objects in Geostationary Orbit

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jiaxin Liu,
  • Feng Yu,
  • Yinghao Wu,
  • Zihan Zhen

摘要

As space activities become more frequent, the safety monitoring of Resident Space Objects (RSOs) in geosynchronous orbit (GEO) is of paramount importance. This paper introduces the MT-DeepLabv3+  methodology, which incorporates spatiotemporal characteristics, aimed at enhancing the detection accuracy of RSOs in GEO. The MT-DeepLabv3+ initiates with a single-frame preliminary check of sequential images via a pretrained DeepLabv3+ network, processing multi-scale information through an encoder-decoder structure while considering edge detail information. To address false positives and missed detections in preliminary results, a novel trajectory post-processing technique is proposed. This method effectively identifies genuine target trajectories, eliminating false positives and compensating for missed detections, significantly improving the detection capability for dim and small targets. Verified on the SpotGEO dataset, our research method demonstrated outstanding efficacy, significantly outperforming current advanced schemes with an accuracy rate of 99.52%, reflected in both F1 score and MSE value. Our approach offers promising application prospects in the detection of dim and small targets at long distances and under low-resolution conditions.