Body motion is essential to our daily lives, and partly reflects our health. With the ubiquitous spread of mobile and wearable devices having built-in micro-electro-mechanical system (MEMS) sensors, acquisition of linear acceleration and angular velocity of personal gait is much easier today. These data can reveal different types of human movement as well as possible abnormalities. This study introduces an approach to collecting and processing human gait data, transforming these data into images and leveraging deep learning technique for the classification task. Features of three-dimensional linear acceleration are re-orientated and decomposed into various frequency components and characteristic waveforms by Zao and Lu method. Then, the summary of these characteristic waveforms produces a feature data set with four classes of human gait motion: walking, jogging, climbing upstairs, and going downstairs. We show promising results with the use of a convolutional neural network and a traditional neural network on the refined feature data. From this study, applications such as tele-monitoring, tele-rehabilitation, and assessing sedentary habits can be implemented to diagnose and intervene in human behavior.