Perceptual Incompleteness in Autonomous Driving: From Information Deficiency to System-Level Propagation
TL;DR
Four improvement strategies are proposed: constructing a joint modeling framework to address the causal coupling among the three types of deficiency, designing probabilistic module interfaces to prevent information attenuation, establishing standardization of cross-module probabilistic semantics, and developing a closed-loop risk-oriented evaluation system.
Abstract
Autonomous driving perception systems face inherent observational incompleteness due to sensor field-of-view constraints and mutual occlusion among targets. The prevailing supervised learning paradigm, which presupposes fully annotated data, fails to provide effective behavioral constraints for models in information-deficient regions. From an information-theoretic perspective, this paper defines perceptual incompleteness as an intrinsic limit imposed jointly by observational geometry and physical mechanisms, and accordingly establishes a three-fold taxonomy comprising spatial, temporal, and causal deficiency. For each type of deficiency, the paper examines its origins, essential characteristics, and existing mitigation strategies. On this basis, the paper examines the systemic limitations of existing approaches from four perspectives: causal coupling, interface attenuation, semantic deviation, and evaluation blind spots. The analysis reveals that the three types of deficiency amplify each other through physical coupling; the perception-to-prediction and prediction-to-planning interfaces filter out uncertainty information through probability binarization and tail branch truncation, respectively; there is no direct mathematical correspondence between perceptual confidence and planning risk probabilities; and current evaluation frameworks systematically exclude information deficiency from the assessment scope. To address these limitations, this paper proposes four improvement strategies: constructing a joint modeling framework to address the causal coupling among the three types of deficiency, designing probabilistic module interfaces to prevent information attenuation, establishing standardization of cross-module probabilistic semantics, and developing a closed-loop risk-oriented evaluation system. The four improvements target training, interface transmission, semantic interpretation, and performance validation respectively, forming a complete solution from algorithm design to system integration to effect validation.