HY-EnYO: A dual-stream YOLO ensemble for deterministic spatial occupancy in seaport container yards

Real-time spatial coordination in high-density seaport container yards is a bottleneck in global supply chains. In these environments, GPS signal denial and severe visual occlusion render traditional theoretical scheduling models ineffective. To resolve this operational gap, this paper introduces HY-EnYO, a novel dual-stream ensemble architecture designed to automate the identification of vacant parking bays using monocular CCTV footage without requiring physical sensor infrastructure. Addressing a fundamental limitation of standard single-stream deep learning detectors, which suffer from severe cross-class feature confusion when mapping massive infrastructure alongside dynamic assets within a unified latent space, the proposed framework decouples visual extraction into parallel sub-networks specialized for yard cranes ($M_c$) and terminal trucks ($M_t$). While this feature extraction remains probabilistic, final occupancy state reasoning is resolved by a deterministic logic engine. This engine projects perspective-aware parking polygons and evaluates Intersection over Union (IoU) overlap against a spatial threshold ($\tau =0.08$), mathematically counteracting camera perspective distortion and sensor jitter. Compared to standard single-stream baselines, HY-EnYO reduces the computational parameter footprint by over 67% (from 36.49 M to 12.02 M), requires a peak VRAM allocation of only 0.10 GB, and achieves an inference latency of 21.82 ms, ensuring edge-device feasibility. Extensive evaluation on real-world CCTV datasets from the Busan seaport demonstrates peak detection accuracies (maP@0.5) of 0.899 for trucks and 0.999 for cranes. Evaluated across a 316-instance test sequence, the deterministic spatial engine achieves a perfect occupancy precision of 1.000 (preventing false-positive routing deadlocks) and an overall system accuracy of 98.0%, maintaining high operational accuracies of 97.0% and 98.0% under simulated adverse weather and low-illumination conditions, respectively. The framework provides a mathematically rigorous, highly efficient spatial matrix, establishing a data-driven foundation for automated seaport traffic coordination.