A technique for detecting scene changes in compressed video streams is proposed which combines multiple modes of information. The proposed technique directly exploits and combines the luminance, chrominance, motion compensation information and the prediction error signal in an MPEG1-coded video stream. By performing minimal decoding of the compressed video stream, the proposed technique results in significant savings in terms of execution time and memory usage. The technique is capable of detecting abrupt scene changes (cuts), gradual scene changes (dissolves), and dominant camera motion in the form of pans and zooms in an MPEG1-coded video stream. Experimental results show that combining multiple modes of information is more effective in detecting cuts and dissolves. The proposed technique is capable of processing video frames in real time and could be used for the rapid generation of key frames for real-time browsing of video streams and for indexing to support content-based access to video libraries.