Visual detection of lintel-occluded doors from a single image

Zhichao Chen, Stanley T. BirchfieldWorkshops @ CVPR 2008 · DOI

Abstract

Reliable door detection is a fundamental requirement for indoor mobile robot navigation and localization. This paper presents a robust, real-time vision algorithm designed to detect doors using a single off-the-shelf webcam. By integrating traditional visual cues with novel geometric features—specifically door concavity and the intensity profile of the gap beneath the door—the proposed system successfully identifies doors even when the top edge (lintel) is occluded. This addresses a common failure mode for robots with low-mounted, downward-tilted cameras.

Introduction

For mobile robots operating in indoor environments, doors serve as critical landmarks. However, detecting them reliably using monocular vision is challenging due to variations in lighting, color, and perspective. A significant hurdle is "lintel occlusion": to avoid floor-level obstacles, robot cameras are often mounted low and angled downwards. As a robot approaches a door, the lintel frequently exits the camera's field of view, causing standard shape-detection algorithms that rely on complete rectangular frames to fail.

Example of a door with an occluded lintel, a common challenge for mobile robot vision.

This research aims to provide a solution that functions across diverse environments, identifying doors whether they are open or closed, textured or plain, and even when they are color-matched to the surrounding walls.

Methodology

The authors approach door detection as a global process, utilizing a Bayesian decision rule implemented via the AdaBoost algorithm. This framework combines multiple "weak" classifiers into a single, highly accurate "strong" classifier. The system evaluates five primary features:

  1. Vertical Line Pairs: An optimized line detection algorithm that merges broken segments caused by hardware like hinges or doorknobs.
  2. Concavity: A novel feature that identifies the "U" shape at bottom corners and the recession shift in the floor-wall boundary, exploiting the fact that most doors are recessed into the wall.
  3. Bottom-Edge Intensity Profile: Another novel cue targeting the gap between the door and the floor, which creates a distinct horizontal line of shadow or light regardless of the door's color.
  4. Color Histograms: The system compares HSV histograms of the candidate door area against the surrounding wall to detect color boundaries.
  5. Texture and Vanishing Points: Texture energy is used to distinguish smooth doors from cluttered backgrounds, while vanishing points ensure the candidate door aligns with the room's perspective.

Experimental Setup and Dataset

To validate the algorithm, a custom database of 309 images was collected across 20 different buildings. The hardware setup consisted of an inexpensive Logitech QuickCam mounted on an ActivMedia Pioneer P3AT mobile robot.

The ActivMedia Pioneer P3AT mobile robot equipped with a laptop and webcam used for data collection and testing.

The dataset was split into 100 images for training AdaBoost weights and 209 images for testing, covering a wide range of floor patterns and lighting conditions.

Results and Discussion

The algorithm demonstrated high performance on the static test set, achieving a 90% detection rate with a false-positive rate of only 0.05 per image. Crucially, the system is computationally efficient, processing images at 5 frames per second on a 1.6 GHz laptop, which is sufficient for real-time navigation.

In live field trials, the robot successfully detected 20 out of 22 doors with 100% accuracy across multiple trials while traversing a 40-meter corridor. The overall detection rate during motion was recorded at 97.7%.

Key insights from the study include:

  • Redundancy via Boosting: No single feature is perfectly reliable, but AdaBoost allows the system to dynamically weigh cues, such as using vertical lines and concavity when color or gap cues are weak.
  • Geometric Robustness: While color and texture vary, the geometric recession of a door into a wall (concavity) is a nearly universal architectural constant, providing a highly robust cue.

Conclusion and Future Work

The proposed method effectively overcomes the lintel occlusion problem by focusing on bottom-of-the-door features and geometric recession. Future work will focus on extending the system to detect open doors, perhaps through motion parallax, and handling doors that are perfectly flush with the wall. Additionally, integrating these visual detections into a full 3D geometric map and implementing automatic wall-color learning could further enhance the robot's navigational capabilities.

Summary updated

BibTeX
@inproceedings{chen2008visual,
  author = {Zhichao Chen and Stanley T. Birchfield},
  title = {{Visual detection of lintel-occluded doors from a single image}},
  booktitle = {Workshops @ CVPR},
  year = {2008},
  doi = {10.1109/cvprw.2008.4563142},
}
Figure