Compression of digital video is the only viable means to transport real-time full-motion video over BISDN/ATM networks. Traffic streams generated by video compressors exhibit complicated patterns which vary from one compression scheme to another. In this paper we investigate the trafficcharacteristics of video streams which are compressed based on the MPEG standard. Our study is based on 23 minutes of video obtained from an entertainment movie. A particular significance of our data is that it contains all types of coded frames, namely: Intra-coded (I), Prediction (P), and Bidirectional (B) MPEG frames. We describe the statistical behavior of the VBR stream using histograms and autocorrelation functions. A procedure is developed to determine the instants of a scene change based on the changes in the size of successive/frames. It is found that the length of a scene can be modeled by a geometric distribution. A model for an MPEG traffic source is developed in which frames are generated according to the compression pattern of the captured video stream. For each frame type, the number of cells per frame is fitted by a lognormal distribution whose parameters are determined by the frame type. The appropriateness and limitations of the model are examined by studying the multiplexing performance of MPEG streams. Simulations of an ATM multiplexer are conducted, in which traffic sources are derived from the measured VBR trace as well as the proposed model. The queueing performance in both cases is found to be relatively close.