In this study, we proposed a continuous stroke phase recognition method with lower-limb inertial signals. The aim of the method was to decrease the time needed and to relieve the burdensome manual configurations in the tasks of human underwater motion recognition. The method automatically segmented the data of a period of time into stroke cycles and three sub-phases (propulsion, glide and recovery). K-nearest neighbor algorithm (k-NN) was used as the classifier to train the segmented data and classify the new data on each sample interval. To validate the proposed recognition method, three elite swimmers were recruited. We also designed an wearable sensing system for human underwater motion sensing with inertial measurement units (IMUs). With only data of 5 stroke cycles for training, the recognizer produced accurate recognition results. The average precision across the phases and the subjects was 93.7% and the average recall was 92.6%. We also investigated the time difference of the key stroke events (stroke phase transitions) between the recognized decisions and the reference ones. The average time difference was 66.2 ms, which accounted for the 4.2% of a single stroke phase. The results of the pilot study proved the feasibility of the new method for human aquatic locomotion assistance tasks. Future efforts will be paid in this new direction for more promising results.