Scholarlink

<p>Object detection on traffic road is crucial for enabling real-time analysis of road conditions and can be applied in intelligent transportation systems. However, in real world, we may encounter cases of low visibility, occlusion and lens contamination, where general methods of object detection usually degrade. In order to address this problem, we propose a Rotary Positional DINO with Siamese Structure (SimRo-DINO) framework, which efficiently overcomes the difficulty associated with object detection under adverse conditions. Specifically, to extract salient detail features and distinguish them from extraneous interference information, we leverage siamese representation learning along with random masking, which is named Mask Siamese Subnetwork, improving the robustness under adverse conditions. Furthermore, to enhance the connection between features scattered by various interferences and capture latent positional information under adverse conditions, we introduce Rotary Position Embedding into Co-DINO framework, an end-to-end detector with the capacity of capturing long-range dependency relationships within images. Extensive experiments have been conduct on UA-DETRAC and our self-built dataset, both under the adverse conditions. The results from our experiments indicate a substantial advancement in mean Average Precision (mAP) of 2.3 and 2.5% on these two datasets, respectively, compared to the Co-DINO baseline. The related codes are publicly available at <a href="https://github.com/xhzhou123/SimRo-DINO">https://github.com/xhzhou123/SimRo-DINO</a>.</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

Simro-dino: a rotary positional DINO with siamese structure for traffic object detection under adverse conditions

  • 期刊论文

  • 发表日期：

    2025年2月11日

查看全文
  • Meng Lei,
  • Xinghan Zhou,
  • Tianju Zhao,
  • Shifan Xu,
  • Liang Zou

摘要

Object detection on traffic road is crucial for enabling real-time analysis of road conditions and can be applied in intelligent transportation systems. However, in real world, we may encounter cases of low visibility, occlusion and lens contamination, where general methods of object detection usually degrade. In order to address this problem, we propose a Rotary Positional DINO with Siamese Structure (SimRo-DINO) framework, which efficiently overcomes the difficulty associated with object detection under adverse conditions. Specifically, to extract salient detail features and distinguish them from extraneous interference information, we leverage siamese representation learning along with random masking, which is named Mask Siamese Subnetwork, improving the robustness under adverse conditions. Furthermore, to enhance the connection between features scattered by various interferences and capture latent positional information under adverse conditions, we introduce Rotary Position Embedding into Co-DINO framework, an end-to-end detector with the capacity of capturing long-range dependency relationships within images. Extensive experiments have been conduct on UA-DETRAC and our self-built dataset, both under the adverse conditions. The results from our experiments indicate a substantial advancement in mean Average Precision (mAP) of 2.3 and 2.5% on these two datasets, respectively, compared to the Co-DINO baseline. The related codes are publicly available at https://github.com/xhzhou123/SimRo-DINO.