A computer vision framework for proactive anomaly detection and risk reduction in airport baggage logistics

This research addresses the challenges in airport baggage handling, focusing on the automated detection of key components-bags, handles, straps-and the identification of damages such as cracks. The proposed system employs the YOLOv8 algorithm for object detection and instance segmentation, trained on a self-generated dataset of 2528 images. For bag accessory detection, the model achieved precision, recall, and F1-scores of 0.92, 0.88, and 0.90 for bags; 0.89, 0.94, and 0.91 for handles; and 0.74, 0.58, and 0.65 for straps, respectively. For damage (crack) detection, YOLOv8’s instance segmentation attained a precision of 0.75, recall of 0.80, F1-score of 0.77, and mean Average Precision (mAP) of 0.76. These results indicate robust detection performance for bags and handles, with scope for improvement in strap and damage detection. A novel aspect of this work is the integration of OpenAI GPT-4 Vision into the baggage inspection pipeline. GPT-4 Vision was employed to perform higher-level reasoning on the detection outputs-such as contextual verification of detected components, natural language description of detected damages, and flagging of anomalies-thereby complementing YOLOv8’s pixel-level predictions with semantic analysis. This hybrid approach enables not only precise localization of components and damages but also contextual interpretation, making the system more adaptable to real-world operational variability. We additionally report deployment-oriented runtime metrics: accessory detection (YOLOv8s) runs at $$\sim$$ 135 FPS (p50 $$\approx$$ 7.4 ms) on an RTX 3090 and $$\sim$$ 110 FPS on a Tesla V100; damage segmentation (YOLOv8s-seg) runs at $$\sim$$ 62 FPS and $$\sim$$ 48 FPS on the same GPUs, respectively, with $$<3$$  GB peak VRAM-comfortably meeting typical 15–30 FPS conveyor camera rates. The results establish new benchmarks for accuracy, reliability, and real-time readiness in baggage inspection. The study highlights the importance of targeted dataset enrichment, statistical validation, and model refinement to address class-specific performance gaps, with significant implications for both research and industrial adoption of intelligent luggage inspection systems.