This paper presents a dual-mode sentinel surveillance architecture for shared vehicles that combines low-power inertial event triggering with low-rate fisheye vision and temporal decision logic. Unlike conventional sentry systems based on continuous vision or motion-only alarms, the proposed approach prioritizes energy efficiency while preserving contextual awareness. We introduce SentinelCAR, a real-world fisheye dataset comprising 18308 images annotated for activity-aware human presence around parked vehicles. A comprehensive benchmark of state-of-the-art detectors (YOLOv8-v12 and RT-DETR variants) shows that YOLOv11m provides the best accuracy-efficiency trade-off, achieving mAP[.5:.95] ≈ 0.90. The selected model is deployed on an NVIDIA Jetson Orin Nano and evaluated under multiple backends and precision formats (FP32, FP16, INT8). Experimental results demonstrate real-time performance of up to 70.4 FPS with (0.92-0.94) J per inference under TensorRT INT8 execution. In addition, an onboard post-processing pipeline based on spatial masking, motion consistency, false-positive suppression zones, and temporal hysteresis significantly improves state stability by reducing spurious activations caused by reflections and static background elements. These results provide a practical foundation for energy-aware, privacy-preserving sentinel surveillance in shared mobility scenarios.
Sentinel Mode Using Sensor Fusion for Improving Security in Shared Vehicles
Sandra Dixe·Duarte Fernandes·Luís De Azevedo Cura·Cátia Loureiro·Tiago Aston·Hugo Peixoto·Ricardo Rodrigues·José Machado·António Silva·Dores Ferreira·Nuno M.C. Da Costa·Beatriz Miranda

