Real-time automatic multilevel color video thresholding using a novel class-variance criterion

Color image segmentation is a crucial preliminary task in robotic vision systems. This paper presents a novel automatic multilevel color thresholding algorithm to address this task efficiently. The proposed algorithm consists of a learning process and a multi-threshold searching process. The learning process learns the color distribution of an input video sequence in HSV color space, and the multi-threshold searching process automatically determines the optimal multiple thresholds to segment all colors-of-interest in the video based on a novel class-variance criterion. For the learning process, a simple and efficient color-distribution learning algorithm operating with a color-pixel extraction method is proposed to learn a color distribution model of all colors-of-interest in the video images, which simplifies the search for optimal thresholds for the colors-of-interest through a conventional multilevel thresholding method. For the multi-threshold searching process, a nonparametric multilevel color thresholding algorithm with an extended within-class variance criterion is proposed to automatically find the optimal upper bound and lower bound threshold values of each color channel. Experimental results validate the performance and computational efficiency of the proposed method by comparing with three existing methods, both visually and quantitatively.