This work presents the development of a markerless optical motion capture system of the front-crawl swimming stroke. The system only uses one underwater camera to record swimming motion in the sagittal plane. The participant in this experiment was a swimmer who is active in the university’s swimming club. The recorded images were then segmented to obtain silhouettes of the participant by a Gaussian Mixture Model. One of the swimming images was employed to generate a human body model that consists of 15 segments. The silhouette and model of the participant were subjected to an image matching process. The shape of the body segment was used as the feature in the image matching. The model was transformed to estimate the pose of the participant. The intraclass correlation coefficient between the results of the developed system and references were evaluated. In general, all body segments, except head and trunk, had a correlation coefficient higher than 0.95. Then, dynamics analysis by SWUM was conducted based on the joint angle acquired by the present work. The simulation implied that the developed system was suitable for daily training of athletes and coaches due to its simplicity and accuracy.