The origin-destination (O-D) matrix of a traffic network is usually estimated from link traffic counts combined with a sample survey. Partially observed vehicle trajectories obtained with vehicle reidentification or automatic vehicle identification techniques such as electronic tags provide a new data source for real-time O-D matrix estimation. However, because of incomplete sampling, accurate estimation of O-D matrices from these data is not trivial. A statistical model was developed for such data, and an unbiased estimator of the O-D matrix was derived based on the method of moments. With further exploitation of the sound statistical model, the bootstrap standard error estimate of the O-D matrix estimator was also developed. The algorithm can be computed quickly and performs well under simulation compared with simpler estimators. Applied to data from vehicles with electronic toll collection tags in the San Francisco Bay Area, the algorithm produces a realistic time series of the hourly O-D matrix. The relationship of the proposed estimator with similar methods in the literature was also studied and extension of the methods to general, more complex networks is discussed.