Lightweight collaborative perception at the edge

This chapter introduces the paradigm of collaborative edge perception, whereby the sensing and DNN-based inference pipelines associated with multiple sensor nodes and edge devices are optimized jointly to overcome both computational and communication resource constraints. Collaborative Edge Perception exploits the fact that multiple sensor nodes often observe the same physical phenomena and/or the same objects, but from different spatial perspectives and/or at different instants of time. Intuitively, such observations provide a degree of redundancy or hints, which can be used to eliminate or reduce unnecessary computation without sacrificing perception accuracy. The chapter describes a core set of techniques (for both RGB and LIDAR sensing data), including DNN state sharing, adaptive attention and dynamic scheduling, that exploit such spatiotemporal correlation to significantly reduce both the communication overheads at the sensor node and the inference overheads at either the edge device or the sensor node. We also describe how such collaboration can help optimize the concurrent execution of multiple perception tasks and outline a set of open problems and prospective approaches.