18 October 2025 in Robotics and Autonomous Systems
6 September 2026 in SSRN Electronic Journal
The deployment bottleneck of autonomous driving perception has shifted from detection accuracy alone to deterministic computational constraints on automotive edge systems. Although recent multi-sensor fusion methods achieve strong benchmark performance, their computational cost and inference latency remain difficult to reconcile with real-time execution on production-grade systems-on-chip (SoCs). This paper presents a deployment-oriented systematic review that frames lightweight sensor fusion as a constrained design problem: reducing latency L, memory footprint M, and energy consumption E while maintaining accuracy A ≥ Amin, inference time T ≤ Tdeadline, and power consumption P ≤ Pbudget. Following the PRISMA protocol, we analyze 59 retained studies and propose a three-dimensional taxonomy spanning fusion level, spatial alignment paradigm, and lightweight optimization. We then synthesize 11 representative algorithms through deployment-oriented evidence fields covering accuracy, reported computational complexity, latency-derived FPS, hardware context, and source category. The review identifies and analyzes four trade-offs thatgovern the deployment envelope, including accuracy–efficiency scaling, modal-ity complementarity, calibration robustness, and safety-performance tension. The analysis shows that knowledge distillation can transfer fusion-relevant