IEEE Network • November/December 200140PEG–4 and H.263 encoded video is expectedto account for large portions of the traffic infuture wireline and wireless networks. To datethe statistical analysis of MPEG–4 and H.263encoded video has received only little attention in the litera-ture. Similarly, there are only few studies that evaluate net-working protocols and resource management schemes withMPEG–4 and H.263 encoded video. This is partly due to alack of sufficiently long frame size traces of MPEG–4 andH.263 encoded videos. In fact, most researchers currently usethe MPEG–1 encodings of Garret [1], Rose [2], Krunz et al.[3], or Feng [4, 5]. These frame size traces give the sizes (inbits or bytes) for each encoded video frame. Networkingresearchers use video frame size traces for video traffic stud-ies [6] and video traffic modeling [7, 8], as well as for thedevelopment and evaluation of protocols and mechanisms forpacket-switched networks [9–14], wireless networks [15], andoptical networks [16]. The cited works are just a small sampleof the hundreds of works that have made use of video tracesover the last couple of years.In this article we report on a new publicly available libraryof frame size traces of long MPEG–4 and H.263 encodedvideos, which we have generated in the TelecommunicationNetworks (TKN) Group at Technical University Berlin. (Thetrace library is available at http://www-tkn.ee.tu-berlin.de/research/trace/trace.html and http://www.eas.asu.edu/trace.)The frame size traces have been generated from MPEG–4and H.263 encodings of over 10 video sequences of 60 mineach. We present a thorough statistical analysis of the framesize traces. We study moments and autocorrelations as well asthe long-range dependence characteristics. We estimate theHurst parameter of the traces with the R/S statistic.This article is structured as follows. We give an overview ofdigital video as well as MPEG–4 and H.263 compression. Wedescribe the generation of the frame size traces. We give anoverview of the encoded video sequences and discuss the cap-turing of the uncompressed video information. We describeour MPEG–4 and H.263 encoding procedures in detail. Weconduct a thorough statistical analysis of the generatedMPEG–4 and H.263 frame size traces. We summarize ourcontributions and, in the Appendix, we review the statisticalmethods used in the analysis of the traces.An Overview of Digital VideoFirst, we give a brief overview of digital video; the interestedreader is referred to [17, 18] for a more detailed discussion. Letus start with an analog video signal generated by an analogvideo camera. The analog video signal consists of a sequence ofvideo frames. The video frames are generated at a fixed framerate (30 frames/s in the National Television Standards Commit-tee, NTSC, format). For each video frame, the video camerascans the frame line by line (with 455 lines in NTSC). To obtaina digital video signal the analog video signal is passed to a digi-tizer. The digitizer samples and quantizes the analog video sig-nal. Each sample corresponds to a picture element (pel). Themost common digital frame formats are Common IntermediateFormat (CIF) with 352 × 288 pels (i.e., 352 pels in the horizon-tal direction and 288 pels in the vertical direction), SourceIntermediate Format (SIF) with 352 × 240 pels, and QuarterCIF (QCIF) with 176 × 144 pels. In all three frame formats,each video frame is divided into three components. These arethe luminance component (Y), and the two chrominance com-ponents: hue (U) and intensity (saturation) (V). Since thehuman eye is less sensitive to the color information than to theluminance information, the chrominance components are sam-pled at a lower resolution. Typically, each chrominance compo-nent is sampled at half the resolution of the luminancecomponent in both the horizontal and vertical directions. (Thisis referred to as 4:1:1 chroma subsampling.) In the QCIF frameformat, for instance, there are 176 × 144 luminance samples, 88× 72 hue samples, and 88 × 72 intensity samples in each videoframe, when 4:1:1 chroma subsampling is used. Finally, eachsample is quantized; typically, 8 bits are used per sample.0890-8044/01/$10.00 © 2001 IEEEMPEG–4 and H.263 Video Traces forNetwork Performance EvaluationFrank H. P. Fitzek, Technical University BerlinMartin Reisslein, Arizona State UniversityAbstractMPEG–4 and H.263 encoded video is expected to account for a large portion of thetraffic in future wireline and wireless networks. However, due to a lack of sufficientlylong frame size traces of MPEG–4 and H.263 encoded videos, most network perfor-mance evaluations currently use MPEG–1 encodings. In this article we present and studya publicly available library of frame size traces of long MPEG–4 and H.263 encoded videos,which we have generated at Technical University Berlin. The frame size traces havebeen generated from MPEG–4 and H.263 encodings of over 10 video sequences 60minutes long each. We conduct a thorough statistical analysis of the traces.MMParts of this work were conducted while Martin Reisslein was with GMDFokus, Berlin, Germany.IEEE Network • November/December 200141As an aside we note that the YUV video format was intro-duced to make color TV signals backward compatible withblack-and-white TV sets, which can only display the lumi-nance (brightness) components. Computer monitors, on theother hand, typically use the RGB video format, which con-tains red, green, and blue components for each pel.Before we discuss the specific features of MPEG-4 andH.263 we briefly outline some of their common aspects. Bothencoding standards employ discrete cosine transform (DCT)[17] to reduce the spatial redundancy in the individual videoframes. Each video frame is divided into macroblocks (MBs).An MB consists of 16 × 16 samples of the luminance compo-nent and the corresponding 8 × 8 samples of the two chromi-nance components. The 16 × 16 samples of the luminancecomponent are divided into four blocks of 8 × 8 samples each.The DCT is applied to each of the six blocks (i.e., four lumi-nance blocks and two chrominance blocks) in the MB. Foreach block the resulting DCT coefficients are quantized usingan 8 × 8 quantization matrix, which contains the quantizationstep size for each DCT coefficient. The quantization matrix isobtained by multiplying a base matrix by a quantizationparameter. This quantization parameter is typically used totune the video