SteerCam: Multi-Camera Edge Perception via Dynamic Joint Steering & Collaboration

Many video surveillance applications utilize networked cameras that are either static or utilize steering to maximize geometric coverage of the sensing field, often neglecting the spatiotemporal scene dynamics. Such applications utilize lightweight DNN models for edge-based execution, which support higher throughput but have significantly lower accuracy compared to deeper DNNs. To overcome these dual drawbacks, we introduce SteerCam, which utilizes Reinforcement-Learning (RL) to jointly adapt networked cameras via: (a) steering, which alters an individual camera’s Field-of-View (FoV) for better visibility, and (b) collaborative inference, where hints from the perspectives of peer cameras are used to selectively enhance each cameras’s inference. We show how using confidence vectors as a salient representation of each camera’s State allows us to overcome the Sim2Real gap and apply an RL model, trained on idealized virtual environments, to diverse real-world scenarios. Using two real-world deployments, we demonstrate that SteerCam with a lightweight SSD detector achieves F1-scores only 5-6% lower than a heavyweight YOLO-v8L model, but with 3-5x higher throughput. SteerCam achieves 1.5-2x higher accuracy than prior steering-only approaches, and 50% lower communication overhead than prior collaboration-only techniques.