Low Latency Wireless Video Over Networks Using Path Diversity John Apostolopolous Wai-tian Tan Mitchell Trott Hewlett-Packard Laboratories Allen Miu MIT Laboratory for Computer Science
Motivation Prevalent infrastructure provides inexpensive connectivity Emergent mobile devices integrated with cameras and interfaces High quality (low-latency) conversational communication over networks (e.g. VoIP, Video conferencing)
Challenges operates in ISM band Interference from other electronic devices (e.g. BT, microwaves) Lossy Environment Access point coverage can be spotty Quality changes over time Signal fading due to multi-path Shadowing due to obstacles and human traffic Contention among exposed and hidden nodes Low Latency requirement ARQ error recovery can add large delays
Our Approach Use error resilient video compression H.264/MPEG-4 AVC Best-effort error recovery (standard ARQ) Distributed AP infrastructure to stream video via multiple access points (exploit path diversity) Use multiple paths simultaneously or switch between them (site selection) as a function of channel characteristics
Preliminary investigation In a path diversity network with mobile clients, How does path diversity affect packet loss characteristics? What is the resulting performance gain for conversational video communication?
Test-bed Setup Sender Mobile Receiver Wired 100Mbps Ethernet b 11Mbps WLAN AP1 AP2 ~25m Ad hoc mode ARQ up to 16 retries Open cubicle area Receiver ~40m (max) Two 360kbps cbr streams 1500 byte UDP packets time-stamped
Diversity Scenarios Conventional single path case Sender Mobile Receiver Wired 100Mbps Ethernet b 11Mbps WLAN AP1 AP2 AP1 only AP2 only
Diversity Scenarios Conventional single path case Sender Mobile Receiver Wired 100Mbps Ethernet b 11Mbps WLAN AP1 AP2 AP1 only AP2 only Balanced split stream (non-adaptive)
Diversity Scenarios Conventional single path case Sender Mobile Receiver Wired 100Mbps Ethernet b 11Mbps WLAN AP1 AP2 AP1 only AP2 only Balanced split stream (non-adaptive) Adaptive, fine-grain site selection (based on loss rate)
Diversity Scenarios Conventional single path case Sender Mobile Receiver Wired 100Mbps Ethernet b 11Mbps WLAN AP1 AP2 AP1 only AP2 only Balanced split stream (non-adaptive) Adaptive, fine-grain site selection (based on loss rate) Oracle (optimal adaptive, can be realized by repetition coding)
Path Diversity Reduces Packet Loss Infinite Avg Packet Loss Rate (%) Delay cutoff (ms) AP1 AP2 Balanced Site Selection Oracle
Path Diversity Reduces Burst Loss Infinite Number of Burst Events Burst event = 2 or more consecutive losses Delay cutoff (ms) AP1 AP2 Balanced Site Selection Oracle
H.263 Video Performance Infinite Mother and Daughter Sequence Delay Cutoff (ms) PSNR (dB) AP1 AP2 Balanced Site Selection PSNR gain = 1.6 – 3.0 dB 1/3 delay reduction
Conclusion All path diversity schemes help reduce burst losses Optimal path diversity drastically reduce loss rate and improves video quality A simple site selection algorithm can effectively increase video quality without increasing bandwidth usage
Complete Trace …
Single Stream from AP … Discarded
Single Stream from AP …
Split stream from AP1 & AP …
… Discarded
Split stream from AP1 & AP … Discarded
Split stream from AP1 & AP … Discarded Re-numbered
Split stream from AP1 & AP …
Fine-grained Site Selection …
… Selected site transmits 95% packet Other site transmits 5% packet for probing Site selection based on error rate of last 300 packets
Oracle … Equivalent to repetition coding
Oracle … Equivalent to repetition coding
Existing Solutions Robust video communication in lossy channel: Error resilient video compression ARQ Delay can be intolerable Head of line blocking FEC coding Can be inefficient Receiver diversity antenna Does not exploit path diversity
Analysis Loss Characteristics Varied delay threshold Average Packet Loss Rate Number of burst events Burst Packet Loss Rate (# Packets lost in burst ) / (Total # packets) H.264/MPEG4 Video Performance for 4 different sequences PSNR N thresh (# of times when PSNR drops below 30dB)
Methodology Walking with receiver in open cubicle area Analyzed packet loss from a 15 minute trace Experiment repeated once after 3 months, showing similar performance results Analyzed the performance of 5 different diversity scenarios from sampling the same trace Difficulty in finding meaningful comparisons between different scenarios