Collaborative video perception leverages pervasively deployed cameras to expand coverage, reduce occlusions, and enhance perception in domains such as transportation, community safety, and robotics. However, analyzing large-scale video streams is data-intensive and strains network and computation resources, especially when only a small region of interest (ROI) is relevant. To address this challenge, we propose PCF (Proactive Communication Framework), with a task-oriented method that incorporates ROI awareness into the detection task objective during training and inference, and an ROI-aware efficient camera selection policy at run time. Evaluations on a public dataset demonstrate up to a 7.5-percentage-point improvement in detection accuracy, while reducing communication cost by 56% without sacrificing performance. These results highlight PCF's potential to enable efficient, ROI-aware perception for battery-powered mobile robots and other resource-constrained edge systems.