Improving Retail Surveillance with Deep Learning: Dataset Optimization and Abnormal Behavior Detection

The increased use of surveillance technology and the need for robust security have fueled research into autonomous, intelligent monitoring systems, particularly in shopping centers where theft and vandalism are common. Deep learning, though promising in anomaly detection, is still hampered by the quality and balance of the training data. In this work, we introduce a deep learning-based anomaly detection system that is trained on an optimized dataset where extensive preprocessing was performed to enhance data quality, eliminate redundancies, and balance class distributions.Our DenseNet-121 and CNN classifier-based system uses this optimized dataset to improve abnormal behavior detection in store surveillance. Experimental results confirm that dataset optimization significantly improves anomaly detection accuracy, reduces false positives, and improves generalizability. Our system achieves an 83% total accuracy, outperforming systems trained on raw, imbalanced data. These findings underscore the importance of data preprocessing as an integral part of deep learning-based surveillance systems.