An Explicit Method for Fast Monocular Depth Recovery in Corridor Environments

Yehao Liu, Ruoyan Xia, Xiaosu Xu, Zijian Wang, Yiqing Ya, Mingze Fan

TL;DR

Researchers have developed a geometric, "explicit" method for monocular depth estimation specifically for corridor environments. By leveraging structural priors and nonlinear optimization instead of heavy deep learning, the system achieves state-of-the-art accuracy while running at real-time speeds (20 FPS) on standard laptop CPUs.

What Problem Does This Paper Address?

Monocular cameras (single-lens) are standard in indoor robotics, but they suffer from a fundamental flaw: they cannot inherently perceive scale or depth.

Current solutions generally fall into two camps, both of which struggle in hallways:

  1. Deep Learning: These models are accurate but computationally "heavy," often requiring expensive GPUs to run in real-time. They also struggle to generalize to new environments not seen during training.
  2. Structure from Motion (SfM): These methods estimate depth by tracking points between moving frames. However, corridors are often "textureless" (plain white walls, uniform lighting), making it nearly impossible for SfM to find reliable points to track.

The authors tackle the "degraded" scenario of long, featureless corridors by using the environment's own geometry—like the parallel lines where walls meet floors—to recover depth without needing a massive neural network.

Methods

The authors propose an explicit method—meaning it relies on mathematical formulas and geometric principles rather than "black box" neural network weights. The system operates in two main threads:

  • Edge Extraction Thread: Uses a Hough Transform to identify the dominant lines in a corridor (the ground-wall junctions). It filters these lines based on "scene priors" (e.g., the expected angle and length of a hallway edge).
  • Depth Recovery Thread:
    • Virtual Camera Assumption: The system projects the real image onto a "virtual camera" imaging plane. This standardizes the view, making the hallway perfectly symmetrical and the optical axis point straight ahead.
  • Nonlinear Optimization: By minimizing the "geometric residual" (the mathematical difference between where edges are and where they should be in a symmetrical hallway), the system solves for the camera’s pitch angle and the actual depth of the points.
  • Depth Plane Construction: Once the edges are mapped in 3D, the system constructs "depth planes" and uses interpolation to fill in the rest of the pixels, creating a dense depth map.

Datasets

The authors introduced a new benchmark called Corr_EH_z.

  • Content: 135 images across 9 different types of corridors.
  • Variability: Includes different lengths (up to 100m), widths (2–4m), and lighting conditions.
  • Subsets: It is divided into Cord_Exx_z (simple, clear corridors) and Cord_Hxx_z (complex corridors with more visual noise).
  • Comparison: The method was also evaluated against the ADABINS model, a high-ranking deep learning approach on the KITTI and NYU datasets.

Key Insights and Ideas

  • Geometry as a Shortcut: In structured environments like corridors, the "rules" of perspective are so strong that you don't need a billion-parameter model to guess depth. You can calculate it if you find the right lines.
  • The Virtual Camera is a Great Equalizer: By transforming images into a standardized virtual space, the algorithm becomes robust to different camera mounting heights and tilts, which usually break simpler geometric models.
  • CPU-Friendly Robotics: Most small robots don't have the power budget for a high-end NVIDIA GPU. This method proves that for specific, common environments (like warehouses or office buildings), we can achieve high performance on a standard mobile CPU.

Key Results

  • Speed: On a mid-range laptop (Intel i5-7300HQ), the method achieved an average execution time of 0.048s per frame (~20 FPS). On a high-end PC, it took just 0.0097s, making it more than 5x faster than the deep-learning-based ADABINS.
  • Accuracy: In the 0–5m range, the method achieved an Absolute Relative Error (AbsRel) of less than 7.1%.
  • Width Estimation: The system was remarkably accurate at estimating the physical width of corridors, with a mean relative error of only 2.21%.

Limitations and Future Work

While highly effective, the method has clear boundaries:

  • Structural Dependence: It requires "closed and straight" corridors. It cannot currently handle curved hallways or open-concept office spaces.
  • Flat Ground Assumption: The math assumes the floor is perfectly level. If there is a ramp or a slope, the depth estimation accuracy drops.
  • Long-Distance Drift: Accuracy decreases at farther distances due to perspective geometry (small pixel errors at the horizon lead to large meter errors in 3D).

Future Work: The authors plan to replace straight-line detection with curve detection to handle non-linear hallways and integrate a lightweight neural network specifically for semantic segmentation to help the geometric solver ignore dynamic obstacles like people.

Summary updated

Abstract

Monocular cameras are extensively employed in indoor robotics, but their performance is limited in visual odometry, depth estimation, and related applications due to the absence of scale information.Depth estimation refers to the process of estimating a dense depth map from the corresponding input image, existing researchers mostly address this issue through deep learning-based approaches, yet their inference speed is slow, leading to poor real-time capabilities. To tackle this challenge, we propose an explicit method for rapid monocular depth recovery specifically designed for corridor environments, leveraging the principles of nonlinear optimization. We adopt the virtual camera assumption to make full use of the prior geometric features of the scene. The depth estimation problem is transformed into an optimization problem by minimizing the geometric residual. Furthermore, a novel depth plane construction technique is introduced to categorize spatial points based on their possible depths, facilitating swift depth estimation in enclosed structural scenarios, such as corridors. We also propose a new corridor dataset, named Corr\_EH\_z, which contains images as captured by the UGV camera of a variety of corridors. An exhaustive set of experiments in different corridors reveal the efficacy of the proposed algorithm.
BibTeX
@misc{liu2023explicit,
  author = {Yehao Liu and Ruoyan Xia and Xiaosu Xu and Zijian Wang and Yiqing Ya and Mingze Fan},
  title = {{An Explicit Method for Fast Monocular Depth Recovery in Corridor Environments}},
  year = {2023},
  eprint = {2309.07408},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2309.07408},
}
Figure