An algorithm for automatic detection of landmarks on three-dimensional faces was proposed by using a feature fusion based constrained local model. A classifier based on depth information and a classifier based on local shape information of three-dimensional meshes were trained for each landmark. The responses of two classifiers were merged and the regularized landmark mean shift algorithm was applied on fitting for the localization of landmarks. Traversing candidates of each landmark is usually necessary in three-dimensional facial landmark localization to generate candidate combinations. The problem of time overhead for nested loops that increases rapidly by using model fitting instead of exhaustive search was solved. The approach was evaluated based on three-dimensional face databases: FRGC v2.0 and Bosphorus. The mean error of every landmark in the FRGC v2.0 is between 2.48 mm to 4.12 mm. The overall detection success rate is 97.3%, among which 97.6% for neutral expression, 97.4% for mild, 95.5% for extreme. On the Bosphorus database, the success rate of 94%, 95% and 89% was respectively achieved under three different poses. The experimental results show that the proposed approach is comparable to state-of-the-art methods in terms of its accuracy, and good robustness is achieved against expression and small pose variation.