Scholarlink

<p>Visual Simultaneous Localization and Mapping (V-SLAM) systems are widely used in autonomous driving and augmented/virtual reality (AR/VR) applications. However, in real-world street environments, images captured by V-SLAM often contain privacy-sensitive content such as faces and license plates, leading to increasing privacy concerns. To address this, we propose a privacy-preserving preprocessing method for V-SLAM and present CIPD-YOLO, a lightweight and accurate model for detecting faces and license plates in real time. CIPD-YOLO reduces model parameters by 31.9% and improves detection accuracy by 12% compared to YOLOv11, significantly enhancing both performance and computational efficiency. The model integrates the Convolutional Block Attention Module (CBAM) to enhance spatial and channel-wise feature representations, while the Involution block further refines feature interaction. We introduce a novel Privacy Object Detection Head (PODH) to enhance detection precision, while the original YOLOv11 prediction head (P5) is removed to increase computational efficiency. Additionally, standard upsampling modules are replaced with Dysample modules to reduce model complexity. To support evaluation, we construct a dedicated dataset, FAPKOR, based on KITTI and Oxford RobotCar, focused on privacy-sensitive object detection. Experimental results show that, when integrated with the proposed preprocessing method, the V-SLAM system maintains high localization accuracy while effectively protecting user privacy in dynamic street scenes. Crucially, we demonstrate that the proposed preprocessing method has no significant impact on SLAM pose estimation accuracy, ensuring that ATE/RPE values remain stable when compared to the original unprocessed input.</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

Privacy-preserving preprocessing for Visual SLAM with an efficient and lightweight object detection model: CIPD-YOLO

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • XiaoLong Ma,
  • ChuHua Huang,
  • MingXu Yang,
  • ShengJin Hou

摘要

Visual Simultaneous Localization and Mapping (V-SLAM) systems are widely used in autonomous driving and augmented/virtual reality (AR/VR) applications. However, in real-world street environments, images captured by V-SLAM often contain privacy-sensitive content such as faces and license plates, leading to increasing privacy concerns. To address this, we propose a privacy-preserving preprocessing method for V-SLAM and present CIPD-YOLO, a lightweight and accurate model for detecting faces and license plates in real time. CIPD-YOLO reduces model parameters by 31.9% and improves detection accuracy by 12% compared to YOLOv11, significantly enhancing both performance and computational efficiency. The model integrates the Convolutional Block Attention Module (CBAM) to enhance spatial and channel-wise feature representations, while the Involution block further refines feature interaction. We introduce a novel Privacy Object Detection Head (PODH) to enhance detection precision, while the original YOLOv11 prediction head (P5) is removed to increase computational efficiency. Additionally, standard upsampling modules are replaced with Dysample modules to reduce model complexity. To support evaluation, we construct a dedicated dataset, FAPKOR, based on KITTI and Oxford RobotCar, focused on privacy-sensitive object detection. Experimental results show that, when integrated with the proposed preprocessing method, the V-SLAM system maintains high localization accuracy while effectively protecting user privacy in dynamic street scenes. Crucially, we demonstrate that the proposed preprocessing method has no significant impact on SLAM pose estimation accuracy, ensuring that ATE/RPE values remain stable when compared to the original unprocessed input.