Adaptive Selective Forwarding Unit for Web Real-Time Communication (WebRTC) Video Conferencing

Background Real-time video conferencing requires maintaining low latency and high video quality even in networks with limited resources. Selective Forwarding Units (SFUs) are widely used in multiparty systems because they are scalable and efficient. However, conventional SFUs forward media streams without adaptation to instantaneous network conditions, which can result in frequent freezes and degraded user experience. Objective This study evaluates an adaptive SFU architecture that combines congestion-control-based bitrate adaptation, real-time quantization parameter control, and keyframe buffering to deliver high-quality video over fluctuating networks. Methods The proposed adaptive SFU was implemented in C++ and compared against a baseline SFU and an Multipoint Control Unit (MCU)-based conferencing tool. Three Web Real-Time Communication (WebRTC) clients (Windows 11, Intel Core i7, Chrome 124, and Android Pixel 7) initiated 720p video calls. Network conditions were emulated using Linux tc/netem with round-trip times of 50/150/300 ms, packet loss of 0/1/5/10%, and a bandwidth sawtooth pattern cycling between 1 and 5 Mbps. Each scenario was run for 5-minute sessions with n = 10 repetitions. We recorded frame rate (FPS), freeze ratio (fraction of time with inter-frame gap > 330 ms), end-to-end delay, peak signal-to-noise ratio (PSNR), and Netflix Video Multimethod Assessment Fusion (VMAF) scores (computed with FFmpeg v7.1). Two-tailed t-tests and bootstrap-based 95% confidence intervals were used for statistical analysis. Results Under demanding network conditions, the adaptive SFU outperformed both the baseline SFU and the MCU. It maintained higher frame rates (32.1 ± 3.2 vs 28.4 ± 2.9 FPS), lower end-to-end latency (~90 ms vs ~100 ms), and substantially improved objective video quality: PSNR increased from 35.2 ± 1.1 to 40.1 ± 1.0 dB and VMAF from 60.3 ± 4.5 to 70.2 ± 3.8. The adaptive SFU also reduced the video freeze ratio from about 40% with the MCU to around 5% by combining adaptive bitrate reduction and on-demand keyframe prioritization. All differences were statistically significant (p < 0.01). Conclusions By combining WebRTC's built-in congestion control with codec-level tuning and keyframe caching, the adaptive SFU mitigates packet loss and jitter and yields less disrupted video than a standard SFU or MCU. Each mechanism (Google Congestion Control [GCC]-driven bitrate adaptation, quantization control, and keyframe replay) contributes incrementally to the observed performance gains. The main limitations are the focus on VP8, evaluation on a modest number of clients in a controlled testbed, and the additional processing overhead on the SFU. Future work will include scaling experiments with larger conferences, support for VP9/AV1 and scalable video coding, and user studies to validate subjective quality.