CadenceNet: Augmenting the OverLoCK Architecture with Temporal-Adaptive Multi-Spectral Contextual Learning for Enhanced Traffic Signal Perception

With the progressive evolution of autonomous driving technologies, the demand for robust and precise traffic signal perception emerges critically to ensure vehicular operational safety. Nevertheless, contemporary deep learning methodologies manifest deficiencies in spectral feature extraction and integration of contextual information, particularly under circumstances involving diminutive or occluded targets. To address such limitations, a CadenceNet architecture founded upon the OverLoCK framework is introduced, adopting temporal-adaptive multi-spectral contextual learning to augment precision and resilience of traffic signal detection. The CadenceNet encompasses three primary modules: the Dichotomous Spectral Orchestration Paradigm (DSOP) facilitates refined discrimination of detailed features via spectral dynamic fusion; the Ephemeral Computational Diversification (ECD) module advances generalization capability through computational graph expansion during training, subsequently optimizing inference efficiency via structural re-parameterization; the Contextual Inception Core (CIC) synthesizes multi-scale contextual priors, subsequently providing semantically coherent outputs to the FocusNet module within OverLoCK, thereby fortifying dynamic convolution kernel representations. Empirical evaluations conducted on the traffic signal datasets (LISA and Bosch Small Traffic Lights) demonstrate achievement of a mean Average Precision (mAP@0.5) of 98.1% and $\mathrm{mAP} \text{@} 0.5: 0.95$ of 78.4%, thus substantiating enhanced detection efficacy for small-scale targets and affirming the feasibility for pragmatic application deployments.