Deep Learning and Intelligent Perception

CLC number: TP368.2

OP-SLAM: An RGB-D SLAMMOT Method Leveraging the Constraints of Object Planar Features

OP-SLAM：一种融合物体平面特征约束的RGB-D SLAMMOT方法

¹ GNSS Research Center, Wuhan University, Wuhan 430072, Hubei, China (武汉大学　卫星导航定位技术研究中心，湖北　武汉　430072)
² School of Geodesy and Geomatics, Wuhan University, Wuhan 430072, Hubei, China (武汉大学　测绘学院，湖北　武汉　430072)
³ School of Electronic Information, Wuhan University, Wuhan 430072, Hubei, China (武汉大学　电子信息学院，湖北　武汉　430072)
⁴ Artificial Intelligence Institute, Wuhan University, Wuhan 430072, Hubei, China (武汉大学　人工智能研究院，湖北　武汉　430072)

^† Corresponding author. E-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.

Received: 10 October 2025

Abstract

By integrating self-localization, environment mapping, and dynamic object tracking into a unified framework, visual simultaneous localization and mapping with multiple object tracking (SLAMMOT) enhances decision-making and interaction capabilities in applications such as autonomous driving, robotic navigation, and augmented reality. While numerous outstanding visual SLAMMOT methods have been proposed, the majority rely only on point features, overlooking the abundant and stable planar features in artificial objects that can provide valuable constraints. To address this limitation, we propose OP (object planar) -SLAM, an RGB-D SLAMMOT system that leverages planar features to improve object pose estimation and reconstruction accuracy. Specifically, we introduce an accurate object planar feature extraction and association method using normal images, alongside a novel object bundle adjustment framework that incorporates planar constraints for enhanced optimization. The proposed system is evaluated on both synthetic and public real-world datasets, including Oxford multimotion dataset (OMD) and KITTI tracking dataset. Especially on the OMD, where planar features are prominent, our method improves object pose estimation accuracy by approximately 60%. Extensive experiments demonstrate its effectiveness in enhancing object pose estimation and reconstruction, achieving notable performance compared with existing methods. Furthermore, OP-SLAM runs in real time, making it suitable for practical robots and augmented reality applications.

摘要

视觉SLAMMOT(同时定位与建图与多目标跟踪)将自定位、环境建图和动态物体跟踪集成于统一框架，在自动驾驶、机器人导航和增强现实等应用中提升决策与交互能力。尽管已有多种优秀的视觉SLAMMOT方法，但它们大多数仅依赖点特征，忽略了人工物体中丰富且稳定的平面特征，而这些特征可提供宝贵的约束信息。为解决这一局限，本文提出OP-SLAM，一种利用平面特征提升物体位姿估计与重建精度的RGB-D SLAMMOT系统。具体而言，我们基于法向图像提出了精确的物体平面特征提取与关联方法，并设计了一种融合平面约束的新型物体捆集调整框架以增强优化效果，所提出的方法在合成数据集及公开真实数据集(包括OMD和KITTI tracking)上进行了验证。大量实验结果表明，该方法在物体位姿估计与重建方面均取得了显著效果，并优于现有方法。尤其是在平面特征显著的OMD数据集上，我们的方法将物体位姿估计精度提升约60%。此外，OP-SLAM可实时运行，适用于实际机器人和增强现实应用。