Spider: Next generation Live Video Analytics over Millimeter-Wave Networks

MSR-TR-2020-17 |

Published by Microsoft

Massive video camera networks are now driving innovation in smart retail stores, road traffic monitoring, and security applications, and realizing live video analytics over these networks is an important challenge that Wi-Fi and cellular networks alone cannot solve. Spider is the first live video analytics network system to use a multi-hop, millimeter-wave (mmWave) wireless relay network to realize such a design. To mitigate mmWave link blockage, Spider integrates a separate low-latency Wi-Fi control plane with the mmWave relay data plane, allowing agile re-routing without suffering the penalty of mmWave beam searching for the new route. With the objective of maximizing video analytics accuracy (rather than simply maximizing data throughput), Spider proposes a novel, scalable flow planning algorithm that operates over hundreds of cameras to simultaneously calculate network routes, load-balance traffic, and allocate video bit rates to each camera. We implement Spider in a mmWave camera network testbed, comparing its object, text, and face detection recall performance against the state-of-the-art wireless video analytics system. Under different video analytic tasks, Spider improves recall by 41.5%—52.9% on average. Further experiments demonstrate Spider’s high scalability and gradual degradation under node and link failures, with a 57% reduction in average failure recovery time.