Detección de anomalías visuales en tiempo real en el automovilismo de alta velocidad: una arquitectura híbrida de...

At 300 km/h, an end-to-end vision delay of 100 ms corresponds to 8.3 m of unobserved travel; therefore, real-time anomaly monitoring must balance sensitivity with strict tail-latency constraints at the edge. We propose a hybrid cache-retrieval inference architecture for visual anomaly detection in high-speed motorsport that exploits lap-to-lap spatiotemporal redundancy while reserving local similarity retrieval for genuinely uncertain events. The system combines a hierarchical visual encoder (a lightweight backbone with selective refinement via a Nested U-Net for texture-level cues) and an uncertainty-driven router that selects between two memory pathways: (i) a static cache of precomputed scene embeddings for track/background context and (ii) local similarity retrieval over historical telemetry-vision patterns to ground ambiguous frames, improve interpretability, and stabilize decisions under high uncertainty. Routing is governed by an entropy signal computed from prediction and embedding uncertainty: low-entropy frames follow a cache-first path, whereas high-entropy frames trigger retrieval and refinement to preserve decision stability without sacrificing latency. On a high-fidelity closed-circuit benchmark with synchronized onboard video and telemetry and controlled anomaly injections (tire degradation, suspension chatter, and illumination shifts), the proposed approach reduces mean end-to-end latency to 21.7 ms versus 48.6 ms for a retrieval-only baseline (55.3% reduction) while achieving Macro-F1 = 0.89 at safety-oriented operating points. The framework is designed for passive monitoring and decision support, producing advisory outputs without actuating ECU control strategies.