A number of obstacle detection systems are now in operation at level crossings to prevent collisions between trains and large vehicles. However, these are not suitable for detecting people on two-wheeled vehicles or pedestrians, who are also frequently involved in accidents at level crossings. To address this problem, we have been developing a new obstacle detection system that utilizes stereo cameras and image processing technology. The stereo camera system is more effective than a single camera system because it measures the three-dimensional shape of obstacles on crossings. This paper describes the development of the new obstacle detection system and its field test results.