Dopamine depletion in Parkinson's increases directed but not random exploration.

Object detection in intelligent transportation systems remains a cornerstone for ensuring safe autonomous driving. While visible-light detectors have achieved remarkable performance [,], they often fail in complex traffic environments due to illumination-induced feature collapse, glare, or low-light conditions [,]. Incorporating thermal (T) information has emerged as a robust approach because infrared sensors capture heat signatures that are largely invariant to illumination changes []. Nevertheless, simple modality fusion does not guarantee reliable detection, as each sensor remains susceptible to specific failure modes in unconstrained traffic scenarios, which are further compounded by remaining difficulties such as modality inconsistency, extreme illumination variations, and adverse weather conditions [–]. Extensive efforts have been devoted to pixel-level and feature-level alignment to alleviate the effects of inherent thermal noise and cross-modality feature misalignment [–], yet existing RGB-T detection frameworks continue to struggle with the effective fusion of complementary RGB and thermal information, resulting in suboptimal weak signal recovery, particularly for distant or small objects [–]. When local discriminative cues are annihilated by sensor degradation or environmental noise [,], conventional architectures suffer from high miss rates [–]. This deficiency largely arises from the absence of mechanisms for contextual verification: the capacity to assess the plausibility of candidate detections using surrounding spatial and appearance relationships. Neuroscientific insights into visual processing suggest that robust perception relies on the integration of local and global information [,]. In particular, the Gestalt completion principle [] indicates that human vision reconstructs missing local information by leveraging relational dependencies across the visual field [,]. Recent multi-modal learning approaches [,] have advanced detection performance by integrating complementary information, but the recovery of weak signals from degraded small objects remains a persistent bottleneck []. We argue that such degraded local cues can be effectively compensated by explicitly modeling candidate detections as relational graphs, capturing both appearance and topological dependencies with their surrounding context [,].