Scholarlink

<p>To address the need for visual junction recognition by resource-constrained Micro Air Vehicles (MAVs) operating in indoor structured environments such as smart warehouses, this paper proposes a lightweight visual junction recognition method to improve the balance between recognition robustness and deployment efficiency in topological navigation. Existing approaches mainly follow two routes: handcrafted features with traditional classifiers and lightweight deep models. The former usually offers low inference cost, but its robustness is limited under stain contamination, partial occlusion, and illumination variation. The latter can achieve higher recognition accuracy, but its continuous inference cost and deployment overhead are relatively high. To this end, a complementary feature-fusion framework based on Foreground Pixel Density (FPD) and Grid-HOG features is developed. By combining feature normalization, group-wise weighting, and Support Vector Machine (SVM) classification, the proposed method integrates global structural distribution information and local gradient-contour information in a controllable manner, thereby improving recognition stability in degraded scenes. Experiments on five datasets show that the proposed method achieves Top-1 accuracy ranging from 95.11% to 98.23% under various degradation and lighting conditions, while maintaining a CPU inference time of approximately 5 ms per image. On the appearance-degradation dataset with stain contamination and partial occlusion, the proposed method achieves 95.11% Top-1 accuracy, clearly outperforming conventional handcrafted-feature baselines, which obtain only 26.67%–58.67%. Meanwhile, compared with lightweight deep models, including ShuffleNetV2_0.5×, MobileNetV3_S, and EfficientNet_B0, the proposed method achieves approximately 3.48, 3.75, and 9.16 times CPU inference speedups, respectively. These results demonstrate its potential for lightweight edge-side deployment in MAV topological navigation.</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

Lightweight Visual Junction Recognition for Topological MAV Navigation Using Complementary Feature Fusion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Yongmei Dou,
  • Mohamad Haniff Junos

摘要

To address the need for visual junction recognition by resource-constrained Micro Air Vehicles (MAVs) operating in indoor structured environments such as smart warehouses, this paper proposes a lightweight visual junction recognition method to improve the balance between recognition robustness and deployment efficiency in topological navigation. Existing approaches mainly follow two routes: handcrafted features with traditional classifiers and lightweight deep models. The former usually offers low inference cost, but its robustness is limited under stain contamination, partial occlusion, and illumination variation. The latter can achieve higher recognition accuracy, but its continuous inference cost and deployment overhead are relatively high. To this end, a complementary feature-fusion framework based on Foreground Pixel Density (FPD) and Grid-HOG features is developed. By combining feature normalization, group-wise weighting, and Support Vector Machine (SVM) classification, the proposed method integrates global structural distribution information and local gradient-contour information in a controllable manner, thereby improving recognition stability in degraded scenes. Experiments on five datasets show that the proposed method achieves Top-1 accuracy ranging from 95.11% to 98.23% under various degradation and lighting conditions, while maintaining a CPU inference time of approximately 5 ms per image. On the appearance-degradation dataset with stain contamination and partial occlusion, the proposed method achieves 95.11% Top-1 accuracy, clearly outperforming conventional handcrafted-feature baselines, which obtain only 26.67%–58.67%. Meanwhile, compared with lightweight deep models, including ShuffleNetV2_0.5×, MobileNetV3_S, and EfficientNet_B0, the proposed method achieves approximately 3.48, 3.75, and 9.16 times CPU inference speedups, respectively. These results demonstrate its potential for lightweight edge-side deployment in MAV topological navigation.