A human color feature extraction method based on the improved Mean shift algorithm and kernel density estimation was proposed with respect to the human feature extraction in the video surveillance system with multi-cameras. It first gets suitable region partition of each person through the improved Mean shift algorithm. Afterwards, kernel density estimation is done to each region’s pixels to attain the color density function. Then, the color feature extraction of each person is realized. The proposed method could attain accurate color model because it automatically gets reasonable region division not by fixed region division fashion, but by human appearance color distribution. Experimental results show the method’s feasibility and robustness. It lays foundation of the human tracking with multi-cameras.