Abstract:
"Shoplifting poses a significant financial and security burden on retailers. Existing detection methods often rely on human vigilance, prone to fatigue and bias, or costly sensor systems. Additionally, AI-based approaches often struggle with adaptability and generalization across diverse retail settings. This paper proposes a novel, adaptable AI-powered system for automated shoplifting detection in retail environments.
The system leverages a combined deep learning approach, utilizing a 2D Convolutional Neural Network (CNN) for spatial information extraction and a Long Short-Term Memory (LSTM) network for temporal dynamics. This combined model enables accurate detection while remaining adaptable to various settings. Furthermore, a semi-supervised active learning approach addresses data scarcity by utilizing unlabelled data effectively. Data augmentation and active learning strategies further enhance generalizability and robustness. The system operates alongside a web application for authorized personnel to review flagged videos and verify potential shoplifting incidents.
Results on a video dataset demonstrate promising performance. Using a confusion matrix for the two-class classification (normal vs. shoplifting), achieved a precision of 0.81 and 0.73 for normal and shoplifting videos, respectively, with recall values of 0.72 and 0.82. The F1-score reached 0.76 and 0.77 for normal and shoplifting videos, respectively. The AUC-ROC score of 0.82 indicates good discrimination ability. These results showcase the potential of the adaptable AI system for improving security and reducing losses in retail environments.
"
