A Scalable and Generalised Deep Learning Framework for Anomaly Detection in Surveillance Videos DOI Creative Commons
Sabah Abdulazeez Jebur, Laith Alzubaidi, Ahmed Saihood

et al.

International Journal of Intelligent Systems, Journal Year: 2025, Volume and Issue: 2025(1)

Published: Jan. 1, 2025

Anomaly detection in videos is challenging due to the complexity, noise, and diverse nature of activities such as violence, shoplifting, vandalism. While deep learning (DL) has shown excellent performance this area, existing approaches have struggled apply DL models across different anomaly tasks without extensive retraining. This repeated retraining time‐consuming, computationally intensive, unfair. To address limitation, a new framework introduced study, consisting three key components: transfer enhance feature generalization, model fusion improve representation, multitask classification generalize classifier multiple training from scratch when task introduced. The framework’s main advantage its ability requiring for each task. Empirical evaluations demonstrate effectiveness, achieving an accuracy 97.99% on RLVS (violence detection), 83.59% UCF dataset (shoplifting 88.37% both datasets using single Additionally, tested unseen dataset, achieved 87.25% 79.39% violence shoplifting datasets, respectively. study also utilises two explainability tools identify potential biases, ensuring robustness fairness. research represents first successful resolution generalization issue detection, marking significant advancement field.

Language: Английский

A Scalable and Generalised Deep Learning Framework for Anomaly Detection in Surveillance Videos DOI Creative Commons
Sabah Abdulazeez Jebur, Laith Alzubaidi, Ahmed Saihood

et al.

International Journal of Intelligent Systems, Journal Year: 2025, Volume and Issue: 2025(1)

Published: Jan. 1, 2025

Anomaly detection in videos is challenging due to the complexity, noise, and diverse nature of activities such as violence, shoplifting, vandalism. While deep learning (DL) has shown excellent performance this area, existing approaches have struggled apply DL models across different anomaly tasks without extensive retraining. This repeated retraining time‐consuming, computationally intensive, unfair. To address limitation, a new framework introduced study, consisting three key components: transfer enhance feature generalization, model fusion improve representation, multitask classification generalize classifier multiple training from scratch when task introduced. The framework’s main advantage its ability requiring for each task. Empirical evaluations demonstrate effectiveness, achieving an accuracy 97.99% on RLVS (violence detection), 83.59% UCF dataset (shoplifting 88.37% both datasets using single Additionally, tested unseen dataset, achieved 87.25% 79.39% violence shoplifting datasets, respectively. study also utilises two explainability tools identify potential biases, ensuring robustness fairness. research represents first successful resolution generalization issue detection, marking significant advancement field.

Language: Английский

Citations

0