Colour Constancy for Visual Surveillance

Aim

A Visual Surveillance system is expected to be operated 24/7, therefore it is inevitable that the illumination of the scene will vary significantly, especially in outdoor environments. A real-time automatic algorithm that can ensure that colour responses are failly constant is required to allow the system to respond reliably to colour-based queries.
The objective of colour constancy is to re-create the scene as it would appear under some known illuminant source (see below)

Colour Variation

Shots of an outdoor scene were captured with rate 1 frame/minute for 20 days.

Scene Samples

To estimate the level of colour variation, we accumulated the colour responses of five pre-defined patches on the image.

Scene View - Locations of 5 patches

The following graph shows the colour responses (Hue, Saturation) for each patch over the observation time period (each colour represents a single patch). The variation of the responses is significant, especially for surfaces of high reflectance. (such as the yellow box at the left or the green metal surface at the right). Therefore, a colour constancy algorithm is necessary to ensure a common colour reference system for long observation periods.

Graph of Colour Responses over 20 days

Colour Constancy Algorithms

We experimented with two well-known algorithms: The Grey World algorithm and the Gamut Mapping
The Grey World algorithm is based on the assumption that the average of spatial reflectances in a scene is achromatic and is applied by a linear transformation.
The Gamut Mapping method attempts to fit a gamut of colour responses of the current frame to the gamut of reference frame.

Automatic Reference Frame Selection

To allow easy initialisation of the system, a method that automatically selects Reference Frame from a video stream was developed. In the HSL colour system, the frame with the highest average of projected saturation magnitude (S) on the plane L=0.5 over all the pixels is selected.
Saturation Projection on the L=0.5 plane
This criterion penalises the chromatic features at either end of the intensity dimension, which can exhibit a high saturation, but relatively little colour information. The selected reference frame is expected to have rich colours.

Selected Reference Frame

Conclusion

According to our evaluation, the Gamut Mapping outperforms the Grey World algorithm. However, the former can be very slow and inappropriate for online processing. Therefore, the Grey World algorithm is suggested for real time applications that quality is not crucial. If quality is the priority, then the Gamup mapping should be used.

Publications

J.R. Renno, D. Makris, T.J. Ellis, G.A. Jones, "Application and Evaluation of Colour Constancy in Visual Surveillance", Joint IEEE International Workshop on Visual Surveillance and Performance Evaluation (VS-PETS 2005), Beijing, China, (2005) abstract download

About this work

This work is part of the EPSRC-funded project REVEAL, conducted by John-Paul Renno, Dimitrios Makris, Tim Ellis and Graeme Jones.