Uncertainty-aware Task Offloading for Multi-vehicle Perception Fusion over Vehicular Edge Computing
Yu-Jen Ku, Sabur Baidya, Sujit Dey
Abstract
Multi-vehicle perception fusion is an emerging advanced vehicular application providing Vehicle Users (VUs) with comprehensive driving assistance. This involves each VU's individual vehicular perception tasks and additional fusion tasks at the end. However, performing these perception fusion applications on some VUs may not be feasible due to the applications' high computing complexity. Vehicular Edge Computing (VEC) server, which is collocated with a Road-side Unit, together with VUs' Vehicular Local Computing (VLC) units can be used to support these perception fusion applications through task offloading. Achieving this edge-collaborated perception fusion with task offloading involves multiple challenges: (i) frequently varying uplink channel conditions, (ii) sequential and parallel task dependencies, and (iii) uncertain task composition induced by unknown object detection outputs in the dynamic driving environment. In this paper, a real-time mechanism is proposed to minimize the end-to-end delay of an edge-collaborated multi-vehicle perception fusion application by addressing these challenges. Based on the real-time channel conditions, VEC and VLC server capacities, and number of detected objects, this algorithm jointly determines the bandwidth allocation, task partitioning, offloading, and execution scheduling for all the involved tasks of the multi-vehicle perception fusion application. Numerical results show that the proposed approach can significantly reduce the end-to-end fusion application latency compared to existing techniques.