A Scalable Platform for Distributed Object Tracking Across a Many-Camera Network

Advances in deep neural networks (DNN) and computer vision (CV) algorithms have made it feasible to extract meaningful insights from large-scale deployments of urban cameras. Tracking an object interest across the camera network near real-time is a canonical problem. However, current tracking platforms two key limitations: 1) They are monolithic, proprietary lack ability rapidly incorporate sophisticated models, 2) less responsive dynamism wide-area computing resources that include edge, fog, cloud abstractions. We address these gaps using Anveshak, runtime platform for composing coordinating distributed applications. It provides domain-specific dataflow programming model intuitively compose application, supporting contemporary CV advances like query fusion re-identification, enabling dynamic scoping network's search space avoid wasted computation. also offer tunable batching data-dropping strategies blocks deployed on respond compute variability. These balance accuracy, its performance, active camera-set size. illustrate concise expressiveness four Our detailed experiments 1000 camera-feeds modest exhibit scalability, quality trade-offs enabled by our tracking, batching, dropping strategies.