Constrained Multi-Objective Bayesian Optimization for Unmanned Ground Vehicle-Oriented Detection Models

Reliable perception of non-NATO armored vehicles is fundamental for Unmanned Ground Vehicle (UGV) operations in safety-critical, time-constrained environments. This study proposes a UGV-oriented framework integrating lightweight You Only Look Once (YOLO) architectures with a constrained multi-objective Bayesian optimization strategy. An original hybrid dataset of 10,640 ground-level images was constructed, featuring tanks, armored personnel carrier (APCs) main battle tank (MBT), self-propelled howitzers, and hard-negatives, excluding aerial views for domain consistency. Quantitative evaluation shows YOLOv9s achieves the highest accuracy reaching 97.33% mAP@50 and 0.8478 mAP@50–95, while maintaining a high recall of 92.11% and the highest Matthews Correlation Coefficient (MCC) score (0.8312). YOLO11s provides the highest sensitivity with a recall of 92.75%, whereas YOLOv5su delivers the lowest latency (13.82 ms) and highest throughput (72.4 FPS), highlighting critical trade-offs between detection accuracy and computational efficiency. To address accuracy-efficiency trade-offs, a Multi-Objective Tree-structured Parzen Estimator (MOTPE) based Bayesian optimization framework yielded Pareto-optimal configurations for Ambush, Reconnaissance, and Balanced mission modes. This approach enables adaptive, hardware-aware model selection while preserving mission-critical detection performance.