Illumination Invariant Background Subtraction for Pan/Tilt Cameras using DoG Responses

In this paper an efficient and robust illumination invariant background subtraction approach for pan/tilt cameras is introduced. In a preprocessing step a panorama-based approach for temporal pixel registration is used to obtain a joint motion independent background model. Hereby, inconsistencies in the background model occur due to alternating illumination, automatic white balancing, AGC, as well as lens vignetting artefacts. During background subtraction such inconsistencies and artefacts lead to segmentation clutter and as a consequence increase false detection rates. To overcome these problems, in this paper an illumination normalization method based on Difference-Of-Gaussian (DoG) band-pass filters is proposed. By preprocessing camera images by the proposed filters lens vignetting artefacts, as well as AGC and global illumination changes are eliminated robustly, and simultaneously local color and intensity information are preserved for accurate motion detection.