The most prominent problems in utilizing the rotoscopy data for human walking animation can be summarized as preservation of the original motion characteristics in the generalization process and constraint satisfaction. Generalization is the process of producing the step of an arbitrary body and step length from the original measured step of one particular subject and step length. If we lose much of the original style in the generalization, it would be meaningless to use the measured data. We present a generalization technique that keeps the original spatial motion characteristics as much as possible. Two types of generalization are considered. One is the anthropometry generalization, which handles the non-uniform segment length ratio differences between the two bodies. The other is the step length generalization, which changes the steps to different step lengths of the same subject. These two generalizations are combined together to generate a step of arbitrary subject and step length. The constraint satisfaction is enforced within the generalization process.