Recognizing human emotions from facial expressions is highly dependent on the quality of the referred facial expression features. Conventional methods often suffer from high computation time and serious influence of environment variations. In this paper, a triangular facial feature extraction method based on a Modified Active Shape Model (MASM) is proposed. This method features considering the interactions of all facial features, escaping from the affection of environment variations as well as noisy facial features, and reducing feature dimensions. MASM adopts the same shape representation and shape training procedures as ASM, but executes a different landmark searching procedure without using the gray level training procedure to avoid the affection from environment variations. Using the feature points extracted by MASM, two methods, one is based on statistical analysis and another one is derived from the genetic algorithm, are proposed to extract an optimal set of triangular facial features for emotion recognition. In the experiments with JAFFE database, a neural network classifier is employed to recognize emotions with those extracted triangular facial features. The experimental results show that based on the statistical analysis 65.1% recognition rate is achieved, and based on the genetic algorithm 70.2% recognition rate is achieved.