Through numerous gait in a lifetime, pedestrians have their own unique gait patterns and many studies have analyzed gait patterns to distinguish the person’s identity Various methods are used for gait analysis to distinguish a person's identity, but existing methods have poor performance in gait prediction in different environment environments. In this study, we used a nonlinear automatic encoder to predict gait 3D kinematics in various environments. Trajectories of joint centers in one gait cycle were represented as a gait data of subject and obtained for 488 subjects by a motion capture system. Each segment’s orientation sequence in three-dimensional gait data was converted into a lower-dimensional vector by two auto-encoder layers. Using gait feature vectors extracted through a trained autoencoder, motion transfer with parameter converting between three different environments performed. This study showed that our method can extract lower-dimensional features of the gait data and can be used for gait motion prediction in different environments.