Development of a Computationally Efficient CNN-Based Deep Learning Model for Anomaly Detection in Retail Surveillance

The retail industry experiences significant financial losses annually due to inventory shrinkage caused by shoplifting. Resource constrained retail environments face significant challenges in detecting anomalous behaviors such as shoplifting due to the high computational demands of advanced anomaly detection systems. These limitations highlight the need for computationally efficient solutions tailored to the specific challenges of such settings. This study introduces a computationally efficient anomaly detection model using a DenseNet121-based convolutional neural network architecture, specifically designed for resource-constrained retail environments. The model was trained and tested on the UCF-Crime dataset, which includes a variety of suspicious and normal activities, to enhance the detection of abnormal behaviors such as shoplifting. By prioritizing reduced computational complexity without sacrificing detection accuracy, our approach achieves a detection accuracy of 68% with an AUC score of 0.65, making it suitable for deployment in resource-constrained settings such as small retail operations. Compared to more complex models like 3D CNNs which achieve $\mathbf{7 5 \%}$ accuracy and an AUC of 0.72 and YOLOv5, which achieves 92% accuracy and an AUC of 0.80, the proposed approach significantly lowers computational demands, making it suitable for real-time deployment in small retail operations. This provides a practical solution for improving security in retail contexts. This research represents a novel step toward accessible and efficient retail surveillance, with the potential to reduce theft-related losses. Future work will focus on addressing class imbalance and improving the detection of complex behaviors.