Uloženo v:
Podrobná bibliografie
Hlavní autor: Yang, Wei
Médium: Recurso digital
Jazyk:Mandarínština
Vydáno: Zenodo 2026
Témata:
On-line přístup:https://doi.org/10.5281/zenodo.20002992
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
_version_ 1866901272542576640
author Yang, Wei
author_facet Yang, Wei
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>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_20002992
institution Zenodo
language cmn
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle 多传感器端到端前融合——基于统一标注的特征级融合方案
Yang, Wei
多传感器融合;1+1<1;允许偏差;特征无冲突;端到端;前融合
&lt;p&gt;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&nbsp;&lt;strong&gt;&ldquo;1+1 &lt; 1&rdquo;&lt;/strong&gt;&nbsp;&ndash; 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)&nbsp;&lt;strong&gt;1+1 &lt; 1&lt;/strong&gt;&nbsp;&ndash; the mathematical inevitability of performance degradation due to hard decision rules; (2)&nbsp;&lt;strong&gt;Allow deviation&lt;/strong&gt;&nbsp;&ndash; abandon perfect spatiotemporal alignment, treat natural deviations as training features to enhance robustness; (3)&nbsp;&lt;strong&gt;Feature non‑conflict&lt;/strong&gt;&nbsp;&ndash; convert multi‑sensor data into stacked feature channels processed by a single neural network, eliminating decision‑level conflicts; (4)&nbsp;&lt;strong&gt;Single‑model output&lt;/strong&gt; &ndash; 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.&lt;/p&gt;
title 多传感器端到端前融合——基于统一标注的特征级融合方案
topic 多传感器融合;1+1<1;允许偏差;特征无冲突;端到端;前融合
url https://doi.org/10.5281/zenodo.20002992