In pedestrian gait classification research, the traditional method based on the micro-electro-mechanical system (MEMS) inertial sensor technique focuses on distinguishing a single pedestrian gait, and ignoring transition gait between two single gaits. It leads to poor classification accuracy of mixed gaits such as walking, running, halting, and even causes loss of time. As a result, positioning error of pedestrian dead reckoning becomes large. This paper analyzes gait characteristics based on kinesiology, and collects raw data of pedestrian using a 9-axis MEMS sensor. 3D inertial sensor parameters are then selected to be applied to the subsequent classification algorithm. Because the Naive Bayes algorithm has low accuracy to distinguish reverse transition gaits, the improved algorithm based on the Naive Bayes algorithm judges continuity of two adjacent windows to realize mixed gait classification. Experimental results show that the proposed mixed pedestrian gait classification method with 3D inertial sensor parameters characterization can distinguish a variety of single gaits and transition gaits from mixed gaits. It can also improve the overall classification accuracy by 14.46% as compared with the method of combining sample entropy and wavelet energy. Therefore, the proposed method has a good theoretical and practical value in gait classification.