Enhancing shoplifting detection precision through quantum-based bayesian optimization techniques

Abstract Shoplifting refers to the unauthorized removal of goods from a store. It remains a critical challenge for retailers, leading to significant financial losses, operational disruptions, and reputational damage. Effectively detecting suspicious behaviours is vital for mitigating these impacts. While traditional countermeasures such as surveillance cameras, security tags, and staff training, which are widely utilized, are limited by their complexity and emerging nature. The YOLO-11 model, the latest object detection model, also faces limitations that reduce real-time inference accuracy. This study introduces an innovative, accurate shoplifting detection system that utilizes pose estimation-based Quantum Bayesian Optimization (QBO) for hyperparameter tuning to identify suspicious activities and enhance YOLO11 more effectively. The system builds upon a meticulously curated dataset with three distinct categories: (1) Low Suspicion (e.g., face coverings during shopping, group entry), (2) High Suspicion (e.g., open bags, cashier distractions, running), and (3) Shoplifting (e.g., concealing items, unauthorized removal of goods, bulk packing of items). The proposed system demonstrates an increment of 6.2%, 10.3%, and 3.8% in true-positive rates for low-suspicion, high-suspicion, and shoplifting types, respectively. The benchmarking enhancements of YOLO11 TPRs for the three classes were as follows: low-suspicion TPR increased by 6.2% from 0.81 to 0.86, high-suspicion TPR increased by 10.3% from 0.87 to 0.96, and shoplifting increased by 3.8% from 0.74 to 0.81. The training cycle is reduced by 60%, from 200 to 80. By integrating advanced object detection models, quantum optimization techniques deliver a scalable, high-performing solution to enhance situational awareness in retail environments.