Soft sensors are becoming more popular in wearables as a means of tracking human body motions due to their high stretchability and easy wearability. However, previous research not only was limited to only certain body parts but also showed problems in both calibration and processing of the sensor signals, which are caused by the high nonlinearity and hysteresis of the soft materials and also by misplacement and displacement of the sensors during motion. Although this problem can be alleviated through redundancy by employing an increased number of sensors, it will lay another burden of heavy processing and power consumption. Therefore, we propose the use of deep learning for human motion sensing, which significantly increases efficiency in the calibration of the soft sensor and estimation of the body motions. First, we make a sensing suit and calibration method for full-body motion tracking. The sensing suit is made of stretchable fabric and contains 20 soft strain sensors distributed on both the upper and the lower extremities. Three athletic motions were tested with a human subject, and the proposed learning-based calibration and mapping method showed higher accuracy than traditional methods that are mainly based on mathematical estimation, such as linear regression. Second, we make a gait motion generating method using only two microfluidic sensors based on semi-supervised deep learning. From the experiment, the proposed method showed a higher performance with smaller calibration dataset comparing to the other methods that are based on supervised deep learning.