Color constancy-based illumination robustness enhancement method for visual SLAM

Abstract

This thesis presents an illumination robustness enhancement method for simulta- neously localization and mapping (SLAM) based on the color constancy estimated by RGB color channels. The core of the method is replacing the intensity image with the illumination invariant intrinsic image generated with the color constancy model whereby resisting the affection of the ambient changing lighting. A series of approaches are developed to incorporate the color constancy model into SLAM. Firstly, a multi-indicator-based approach is developed to improve the distinc- tiveness of individual points on the image to improve the accuracy of the tracking process. In contrast to the single indicator color constancy estimation approach, the proposed method uses three indicators to calculate independent illumination invariant components to keep more color information. A Kalman filter-based optimization is then applied to these three indicators to minimize the affection from the measurement error and enhance the performance under changing illumination. In addition, the gradient magnitude is used to improve the distinctiveness of each image pixel. A series of experiments show the proposed approach effectively increases the tracking accuracy and illumination robustness in both the direct and feature-based SLAM. Secondly, A fast color constancy adjustment approach is introduced to improve the SLAM robustness in the indoor environment. In contrast to the outdoor natural illumination, the multiple lighting sources in the indoor environment bring more error in the color constancy estimation, The proposed approach takes in to account this error and leverages the revised optical flow method to provide a fast color constancy estimation process. In contrast to other estimation methods, the proposed approach transforms the direct search problem to a least square problem and uses an iteration algorithm to improve the efficiency and accuracy. A parallel color constancy estimation process is also introduced to the SLAM to balance the robustness and computational complex. Finally, an hierarchy method is proposed to deal with the problem of diverse order of magnitude. The color constancy model is based on the logarithm operation which produces an intrinsic image that is robust to the illumination change but with

discrepant order of magnitude. As such, it is hard to set a global threshold for filtering operator and calculating the local gradients within same order of magnitude. The proposed method utilizes a hierarchy model to adjust the contrast of the output result therefore obtains more even distance distribution. In a conclusion, the proposed approaches in this thesis successfully incorporate the color constancy model into the illumination robust SLAM with lower compu- tational cost. In contrast to the state-of-the-art illumination robust SLAM methods which reach real-time performance by using GPU, the proposed method can work at 20Hz on an Intel i7 2500Hz CPU. The experiments show that the method presented in this thesis has high robustness in both the long-term and short-term SLAM tasks and the potential to be applied to both the direct and feature-based SLAM.