Colour segmentation for multiple low dynamic range images using boosted cascaded classifiers

This paper proposes a multi-camera system approach to real-time colour segmentation using a cascade of AdaBoost colour classifiers with error-correcting codes (AdaBoost ECC) trained on arbitrary low-dynamic range cameras. As compared to traditional High-Dynamic Range systems that require the consolidation of multiple low-dynamic range (LDR) images to produce a single HDR image or a single tone mapped image, the proposed approach feeds directly on the LDR images, and is therefore less computationally intensive. Furthermore, the proposed approach can be employed without necessitating any spectrometric calibration of the cameras. It treats each chromatic/achromatic channel of the multi-camera system as a feature vector, allowing for a multi-dimensional colour search space over a combination of different cameras and colour spaces, without any theoretical limit to the number and type of cameras. A scene plagued with spatially-varying illumination conditions was used to test the efficacy of the proposed system. A dataset of over 89,000 samples were used for training the AdaBoost classifiers to learn eight colour categories in a matter of minutes. The experiments showed that under these extreme illumination conditions, the classifier using three cameras achieved 93% correct classification compared to less than 75% when using a single camera.