This paper presents a depth-informed approach for detecting abandoned luggage in surveillance environments, combining fine-tuned object detection and dynamic spatial reasoning. A YOLOv11-m model was optimized for person and luggage detection using a custom surveillance dataset captured in real-world conditions. To improve the reliability of abandonment detection in varying perspectives and crowded scenes, the system uses depth maps to dynamically scale the proximity radius between individuals and luggage, accounting for perspective distortion. To address the scarcity of real abandoned luggage events needed for system validation, we evaluated three scenario generation methods: actor-based simulations, generative AI models (Open-AI’s Sora and MiniMax’s Hailuo), and scene augmentation techniques. Our findings highlight the current limitations of generative AI in producing specific security related scenarios and underscore the effectiveness of augmentation techniques. The proposed system demonstrates robust detection performance and contributes a reproducible methodology for rare-case simulation. The research contributes with a specialized surveillance training dataset, a fine-tuned YOLOv11 model for person and luggage detection and an algorithm for abandoned luggage recognition. Also, rare synthetic scenes are generated for robustness evaluation in smart luggage detection surveillance systems along with analysis of generation methods.