Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | Chinese |
| Published: |
Zenodo
2026
|
| Subjects: | |
| Online Access: | https://doi.org/10.5281/zenodo.20002992 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <p>Current autonomous driving multi‑sensor fusion systems generally adopt a post‑fusion (decision‑level) architecture. This paper identifies a fundamental defect of post‑fusion, precisely described as <strong>“1+1 < 1”</strong> – the overall performance after fusion is worse than that of the better single sensor. The paper proposes a complete end‑to‑end early‑fusion solution and distills four core principles: (1) <strong>1+1 < 1</strong> – the mathematical inevitability of performance degradation due to hard decision rules; (2) <strong>Allow deviation</strong> – abandon perfect spatiotemporal alignment, treat natural deviations as training features to enhance robustness; (3) <strong>Feature non‑conflict</strong> – convert multi‑sensor data into stacked feature channels processed by a single neural network, eliminating decision‑level conflicts; (4) <strong>Single‑model output</strong> – a single network directly outputs the final result, reducing latency and error accumulation. From information theory, probability, and engineering practice, the paper demonstrates these principles and provides a unified implementation framework. The proposed scheme theoretically surpasses the performance ceiling of post‑fusion architectures.</p>