To improve the accuracy of serve reception and serve effectiveness in volleyball, it is important to provide players with feedback about ball trajectories for effective serves. However, manual acquisition of the threedimensional coordinates of the served ball is extremely time-consuming (about 30 minutes per serve). Because this hinders the provision of immediate feedback, an automatic and time-efficient system is required. The purpose of this study was to develop an automatic system for acquisition of ball trajectories in volleyball serves using a convolutional neural network. Twenty serves during an official game were recorded with 4 digital video cameras (60 Hz). In this system, two-dimensional coordinates of the ball center on an image were acquired using the convolutional neural network, and then three-dimensional coordinates were calculated. To test the accuracy of the developed system, three-dimensional coordinates were compared with the data obtained by manual digitization. The average root mean squares of differences in the coordinates of the ball center between the developed system and the digitized data were 0.014±0.005 m in the X direction, 0.008±0.003 m in the Y direction, and 0.010±0.003 m in the Z direction. The time required to obtain three-dimensional coordinates using this method was 158.6±21.5 s per serve. These results suggest that the present system facilitates acquisition of ball trajectories with an accuracy almost the same as that for digitization, but in a shorter time.