Rapid and Robust Algorithms for Detecting Colour Targets

Introduction Colour is a distinctive feature that can vastly simplify visual search for objects. However, some form of colour constancy is needed to cope with unknown lighting conditions. Unfortunately, current computer vision colour constancy algorithms generally make restrictive assumptions about illumination conditions and/or require significant computational resources. We propose a computer vision algorithm for the rapid and robust detection and localization of specially designed colour targets. The algorithm draws on properties of colour gradients which are invariant to a wide range of illumination conditions, without needing to implement full colour constancy. It is intended to be used for an environmental labelling system for helping visually impaired people navigate their environment. Since the algorithm needs to run in real-time on an inexpensive handheld PDA/cell phone with colour camera (see Figure 1a), which has limited computational power, algorithmic efficiency is of paramount importance. Our algorithm for detecting colour targets exploits the fact that, although the apparent target colours depend strongly on lighting conditions, colour gradients within them vary much less. Moreover, the spatial and spectral properties of the targets make them easy to distinguish from typical background clutter. Colour targets are located in an image through a series of simple colour gradient tests. Targets may be placed adjacent to signs or barcodes, which can be rapidly read given the location of the target (and conveyed to the user by synthetic speech). The targets are small and can be detected from a distance; the maximum range is proportional to the size of the target and is a few meters from the camera using a 4 cm diameter target. Detection takes 0.12-0.5 sec per image in our prototype system, demonstrating its feasibility for real-time use.