Visual Detection of Lintel-Occluded Doors
from a Single Image

Zhichao Chen and Stan Birchfield

Doors are semantically meaningful objects crucial for indoor navigation and mapping.  We have developed a technique that detects doors in a single image in near real time (approximately 5 fps) from a camera mounted on a mobile robot.  Because the camera is mounted low to the ground, the top (lintel) of the door is often not visible.  To solve the problem in a variety of environments, several standard and novel cues are employed.  Any single cue is unable on its own to detect doors reliably, but these weak classifiers are combined using Adaboost to produce a strong classifier capable of reliable detection.

Algorithm Overview

The basic algorithm is represented below.  Vertical intensity edges are used to detect door candidates, which are then tested using the remaining cues.  

Experimental Results

The algorithm achieves 90% accuracy with 1 false positive for every 20 images, on a database of 209 images collected in 20 different buildings. 


Example images:


Videos:



Video of camera on left



Video of camera on right

Database of 440 door images:  door_images.zip

Publications