Abandoned Bag Detection in Public Areas Using Grounding DINO With Fine-Grained Prompts

Detecting abandoned bags in public areas is essential for ensuring safety and preventing potential threats. This research presents a novel approach for abandoned bag detection using Grounding DINO, a state-of-the-art vision-language model, combined with fine-grained semantic prompts. The method integrates object detection and contextual analysis to identify stationary bags left unattended for specified durations, distinguishing them from routine activities in dynamic settings such as airports, train stations, and shopping malls. The proposed approach is evaluated on the ABODA dataset, achieving an abandoned bag detection recall of 90.91%. Grounding DINO's capability to process textual prompts ensures precise bag identification, while its adaptability supports diverse public environments. The workflow involves detecting bags using fine-grained prompts, tracking their movement across frames through centroid distances between the owner’s and bag’s bounding boxes, and applying temporal and frame-based criteria to confirm abandonment. To enhance reliability, the system incorporates proximity-based owner identification by detecting nearby individuals and analyzing their interactions. Context-aware thresholds adjust detection parameters, ensuring robustness in crowded or complex environments. Furthermore, the results are compared to previous works to evaluate differences in performance and capabilities.