Scholarlink

<p>In-vehicle infrared object detection supports autonomous driving and night surveillance. However, detecting small targets in nighttime infrared images remains challenging due to high noise, low contrast, and poor feature extraction. We propose LCAF-RTDETR (Locality Context Calibration Fusion Real-Time Detection Transformer), a lightweight infrared object detection model with two key modules: a locality sensitive hash (LSH)-based spatial context recalibration module (LCRM) and an adaptive spatial fusion module (ASFM). LCRM efficiently identifies spatially sensitive regions with minimal computational overhead to enhance contextual understanding. ASFM employs adaptive kernel convolution for effective multi-scale feature fusion and spatial adaptability enhancement, particularly benefiting small object detection in complex nighttime scenes. Experimental results on the FLIRV1 dataset demonstrate that LCAF-RTDETR with ResNet-18 backbone achieves 63.1 mAP, outperforming the baseline RTDETR by 4.3% while maintaining real-time inference speed. This improvement, achieved without sacrificing computational efficiency, validates our model’s effectiveness for real-world applications like autonomous driving and surveillance systems.</p>

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

LCAF-RTDETR: real-time infrared object detection with locality-aware feature recalibration and adaptive fusion

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Han Li,
  • Binyu Zhu,
  • Baofu Wu,
  • Xin Zhang,
  • Qianru Sun,
  • Rong Jiang,
  • Hongbo Wang,
  • Yuyu Yin

摘要

In-vehicle infrared object detection supports autonomous driving and night surveillance. However, detecting small targets in nighttime infrared images remains challenging due to high noise, low contrast, and poor feature extraction. We propose LCAF-RTDETR (Locality Context Calibration Fusion Real-Time Detection Transformer), a lightweight infrared object detection model with two key modules: a locality sensitive hash (LSH)-based spatial context recalibration module (LCRM) and an adaptive spatial fusion module (ASFM). LCRM efficiently identifies spatially sensitive regions with minimal computational overhead to enhance contextual understanding. ASFM employs adaptive kernel convolution for effective multi-scale feature fusion and spatial adaptability enhancement, particularly benefiting small object detection in complex nighttime scenes. Experimental results on the FLIRV1 dataset demonstrate that LCAF-RTDETR with ResNet-18 backbone achieves 63.1 mAP, outperforming the baseline RTDETR by 4.3% while maintaining real-time inference speed. This improvement, achieved without sacrificing computational efficiency, validates our model’s effectiveness for real-world applications like autonomous driving and surveillance systems.