Abstract: To deal with the real 3D scene data faced by the mobile robot, a visual saliency detection method based on spatial-spectral mixture analysis is proposed. Firstly, a multichannel feature fusion algorithm is developed to integrate the color and depth information in RGB-D data. Then, the hypercomplex Fourier transform is applied in frequency domain to getting the multi-scale saliency maps. After that, an uneven superpixel segmentation algorithm is used to smooth each obtained saliency map. In this way, the interference of discrete background noises is eliminated and the detection result in frequency domain is improved. Finally, the cellular automata algorithm is adopted to merge the multi-scale visual saliency maps and extract the final saliency region. Plenty of experiments are conducted on the public database to verify the effectiveness of the proposed method in the real complex scene faced by the mobile robot.