Motion enhanced video anomaly detection using masked autoencoder and hybrid loss functions
Mohammed Iqbal Dohan Almurumudhe, Olivér Hornyák · Az Eszterházy Károly Tanárképző Főiskola tudományos közleményei. Tanulmányok a matematikai tudományok köréből/Az Eszterházy Károly Főiskola tudományos közleményei. Tanulmányok a matematikai tudományok köréből/Annales mathematicae et informaticae · 2025
In this paper, a hybrid deep learning framework for video anomaly detection that combines autoencoder-based reconstruction with an advanced anomaly scoring mechanism is proposed. Unlike conventional methods that rely solely on reconstruction loss, our approach integrates motion-based scoring and masked autoencoders to enhance detection accuracy and interpretability. The autoencoder learns to reconstruct normal patterns, while an anomaly scoring function evaluates deviations based on reconstruction errors and motion gradients. This directs attention to dynamic regions and foreground objects, thereby reducing false positives from background variations. To improve robustness, we apply preprocessing techniques, including min-max normalization and data augmentation (random cropping, horizontal flipping, and rotation), ensuring consistency across datasets. The framework is evaluated on widely used benchmark datasets, ShanghaiTech Campus and UCSD Ped2, using precision, recall, ROC-AUC, and confusion matrices as performance metrics. It outperforms traditional reconstruction-based autoencoders and GAN-based models. Furthermore, the hybrid scoring mechanism reduces false positives by 15% compared to standard autoencoder approaches, improving detection reliability. Despite the high accuracy, the method incurs additional computational overhead due to motion gradient calculations and masked reconstructions. However, the trade-off is justified by significant improvements in anomaly detection performance. The results demonstrate that our framework enhances both accuracy and interpretability, making it a viable solution for real-world applications such as surveillance, traffic monitoring, and industrial security.