The Four Core Principles of Multi-Sensor End-to-End Front Fusion —— A Feature-Level Fusion Scheme Based on Unified Annotation

Current multi-sensor fusion systems in autonomous driving generally adopt a post-fusion architecture, and their inherent flaw can be accurately described as 1+1<1 — the overall performance after fusion is lower than that of the optimal single sensor. This paper proposes a complete end-to-end front fusion scheme and summarizes four core principles: (1) The mathematical essence of 1+1<1: the hard decision rules of post-fusion will inevitably lead to performance degradation; (2) The principle of allowing deviations: abandon the perfect spatiotemporal alignment of sensor data, and use natural deviations as training features to enhance robustness; (3) The principle of no feature conflict: convert multi-sensor data into feature channel superposition, which is uniformly processed by the neural network to completely eliminate decision-level "conflicts"; (4) The principle of single-model output: a single network outputs the final result end-to-end, avoiding delays and error accumulation caused by multi-level coupling. This paper demonstrates these four principles one by one from the perspectives of information theory, probability theory and engineering practice, and presents a unified implementation framework. Theoretically, this scheme can break through the performance ceiling of post-fusion.