Coral-pie: A geo-distributed edge-compute solution for space-time vehicle tracking

Abstract

We present a distributed system architecture which is scalable by design for cross-camera vehicle tracking at video ingestion time dubbed Coral-Pie. To meet the latency bounds for timely processing of every frame at each camera, we associate dedicated low-cost computational resource for each camera, which consists of two Raspberry Pi 3B+'s and one Coral Accelerator (EdgeTpu). The end-to-end system generates and stores the tracks in a graph database for easy querying. We use the Cloud-Edge-Device continuum to appropriately place the components of the distributed system architecture. Using the timing profiles of the sub-tasks involved in the continuous processing that needs to happen on every frame in each camera, we map the elements of the processing onto the computational resource associated with each camera. Performance evaluation of the proof-of-concept system is conducted using live streams from five campus cameras. The evaluation includes microbenchmarks as well as application level studies. The controlled experiments using live cameras are augmented with a simulation-based study to show the self-healing property of the system and the system scalability.