Object-Based Reliable Visual Navigation for Mobile Robot

Fan Wang, Chaofan Zhang, Wen Zhang, Cuiyun Fang, Yingwei Xia, Yong Liu, Hao DongSensors 2022 · DOI

TL;DR

This paper introduces a novel object-based visual navigation system for mobile robots that reduces reliance on precise metric maps. By constructing a lightweight topological semantic map and employing an object-guided path planning and trajectory generation strategy, the system achieves more reliable and efficient navigation in complex environments.

What Problem Does This Paper Address?

Autonomous mobile robots are increasingly used in diverse applications, making reliable navigation crucial. While many existing visual navigation methods rely on pre-constructed, high-precision metric maps (like occupancy grids) or require extensive training data for learning-based approaches, these methods struggle in unstructured or dynamic environments. The challenge lies in developing navigation systems that are robust, efficient, and less dependent on perfect environmental models, especially in real-world scenarios where metric maps can be difficult to maintain and scale.

Methods

The proposed approach tackles this by introducing an object-level topological visual navigation method. Key components include:

  • Lightweight Object-Level Topological Semantic Map: Instead of precise metric maps, this method builds a graph-based representation where nodes are objects and edges represent their semantic and topological associations. This reduces the need for detailed 3D geometry.
  • Object-Based Heuristic Graph Search: An A*-like search algorithm is employed on the topological map. It uses semantic-geometric features of detected objects (class, color, 3D position, scale) to guide the search for an optimal and shortest global topological path. Robot-centric relative topological associations are used to provide weights in the absence of global poses.
  • Object-Guided Global Path Segmentation: To mitigate cumulative errors from global paths, the global path is segmented based on active visual perception and object guidance. This breaks down the navigation task into smaller, manageable segments.
  • Bernstein Polynomial-Based Smooth Trajectory Refinement: Each segmented path is then refined into a smooth, dynamically feasible trajectory using Bernstein polynomials. This transforms trajectory generation into a nonlinear planning problem, ensuring smooth and efficient movement.

Datasets

  • Simulation: The Gibson dataset was used for simulation experiments. The authors selected nine scenes from this dataset, excluding those with multiple floors or empty rooms.
  • Real-World: The authors used two typical indoor environments (weakly textured corridors and offices) for real-world experiments. They also mention collecting and adding data for training based on the SunRGBD dataset format for their 3D object detection.

Key Insights and Ideas

  • Shifting from Metric to Semantic Topology: The core idea is that for reliable navigation, understanding the relationships and semantic meaning of objects is often more crucial and scalable than maintaining a precise geometric map of the entire environment.
  • Object Features as Navigation Cues: Leveraging rich semantic-geometric features of objects (beyond just visual appearance) provides robust landmarks for navigation, even under varying illumination or partial occlusion.
  • Hierarchical Navigation: The approach combines global topological path planning with local trajectory generation and segmentation, effectively managing complexity and reducing cumulative errors.
  • Bernstein Polynomials for Smoothness: Using Bernstein polynomials for trajectory generation offers a principled way to ensure smooth, dynamically feasible paths, which is essential for robot motion.

Key Results

  • Simulation: The proposed method achieved significantly higher Success Rate (SR) and Success weighted by inverse Path Length (SPL) compared to state-of-the-art learning-based methods (NTS, ANS, MSMRL) across easy, medium, and hard scenarios. For instance, in the "Hard" scenario, the SR improved from 0.43 to 0.75, and SPL from 0.26 to 0.63.
  • Real-World: The method was compared against OGMADWA (a classical metric-map-based approach). While SRs were comparable, the proposed method consistently achieved higher SPLs, indicating better efficiency. Notably, the method demonstrated robustness in challenging scenarios like dynamic obstacles and long, weakly textured corridors.
  • Robustness: Experimental results showed high robustness of the 3D object detection component to illumination variations, occlusion, and dynamic objects.

Limitations and Future Work

  • Dynamic Objects: The current approach models the environment based on static object-level landmarks. If objects are dynamic and their topological relationships change significantly, the robot risks getting lost. The authors acknowledge this limitation and propose setting up five states for the robot (initialization, waiting, trajectory regeneration, re-planning, moving) to handle such situations.
  • Lightweight Object Detection: The authors plan to use a lightweight object detection network to improve accuracy and reduce computational complexity, enabling deployment on embedded platforms.
  • Large-Scale Navigation: Future work will explore extending the proposed method to more large-scale real-world navigation tasks.
  • Dynamic Relationship Changes: Further improvements are needed to enhance navigation performance in cases of dynamic changes in relationships between objects.

Summary updated

Abstract

Visual navigation is of vital importance for autonomous mobile robots. Most existing practical perception-aware based visual navigation methods generally require prior-constructed precise metric maps, and learning-based methods rely on large training to improve their generality. To improve the reliability of visual navigation, in this paper, we propose a novel object-level topological visual navigation method. Firstly, a lightweight object-level topological semantic map is constructed to release the dependence on the precise metric map, where the semantic associations between objects are stored via graph memory and topological organization is performed. Then, we propose an object-based heuristic graph search method to select the global topological path with the optimal and shortest characteristics. Furthermore, to reduce the global cumulative error, a global path segmentation strategy is proposed to divide the global topological path on the basis of active visual perception and object guidance. Finally, to achieve adaptive smooth trajectory generation, a Bernstein polynomial-based smooth trajectory refinement method is proposed by transforming trajectory generation into a nonlinear planning problem, achieving smooth multi-segment continuous navigation. Experimental results demonstrate the feasibility and efficiency of our method on both simulation and real-world scenarios. The proposed method also obtains better navigation success rate (SR) and success weighted by inverse path length (SPL) than the state-of-the-art methods.
BibTeX
@inproceedings{wang2022object,
  author = {Fan Wang and Chaofan Zhang and Wen Zhang and Cuiyun Fang and Yingwei Xia and Yong Liu and Hao Dong},
  title = {{Object-Based Reliable Visual Navigation for Mobile Robot}},
  booktitle = {Sensors},
  year = {2022},
  doi = {10.3390/s22062387},
}
Figure