Anomaly Detection in Surveillance Videos
In the present society, video surveillance systems are rapidly evolving with intelligent video analytics to improve public safety. With the increasing installation of surveillance cameras in both public and private spaces, there is a growing reliance on continuous monitoring to ensure public safety....
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Year Project / Dissertation / Thesis |
| Published: |
2025
|
| Subjects: | |
| Online Access: | http://eprints.utar.edu.my/7147/1/3E_2003619_FYP_report_%2D_JIA_QI_FOO.pdf http://eprints.utar.edu.my/7147/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In the present society, video surveillance systems are rapidly evolving with intelligent video analytics to improve public safety. With the increasing installation of surveillance cameras in both public and private spaces, there is a growing reliance on continuous monitoring to ensure public safety. However, human-based monitoring is labour-intensive and inefficient. Video anomaly detection (VAD) plays a vital role in modern surveillance systems by
automatically identifying unusual events in video streams. This study focuses on developing a lightweight and efficient VAD framework that supports both binary and multiclass detection. The proposed system, AnomLite combines MobileNetV2, a lightweight Convolutional Neural Network (CNN) for spatial feature extraction, and Long Short-Term Memory (LSTM) for temporal modelling. By leveraging the strengths of MobileNetV2 in extracting efficient spatial features and LSTM in capturing temporal dependencies in video sequences, the model detects anomalous events across various classes. The
system trains on two datasets: UCF-Crime, which contains real-world CCTV footage, and XD-Violence, which includes video content from movies and YouTube. Preprocessing steps are employed to ensure the model performs well under varying data conditions. The evaluation of the proposed model shows
strong performance on the first dataset, achieving an ROC AUC of 0.99 and an average precision of 0.99 on UCF-Crime. The model demonstrates strong performance on another well-known dataset in video anomaly detection, achieving an ROC AUC of 0.98 and an average precision of 0.97 on XD Violence. The model also achieves high accuracy of 94% on UCF-Crime and
93% on XD-Violence, with strong F1 scores across both datasets (F1-Micro 0.93 on UCF-Crime, 0.89 on XD-Violence). The model achieves high per-class accuracy across the UCF-Crime dataset, with 10 out of 14 classes exceeding 0.95 accuracy and several classes, such as Arson, Explosion, Fighting, Shooting, and Vandalism, reaching a perfect accuracy of 1.00, demonstrating the model’s strong and consistent performance in detecting diverse types of anomalies. Moreover, the model performs well on the XD-Violence dataset, with accuracies ranging from 0.79 to 0.95. It shows highest accuracy on Car Accidents (0.95) and strong performance across other classes like Abuse, Riot,
and Fighting, indicating its effectiveness in handling diverse anomalies. Additionally, the model is optimized for inference through quantization. With a reduction of around 70% in model size through model compression techniques
such as quantization, the flexibility of the model is further improved, particularly for low-end devices. These results highlight how deep learning techniques, such as SMOTE, data augmentation, and advanced loss functions like cross-entropy loss, contribute to high accuracy and effective performance
in automating surveillance tasks, even when dealing with highly imbalanced datasets. Data augmentation techniques that simulate real-world conditions enhance the efficiency of anomaly detection systems in practical applications.
Keywords: Video anomaly detection, deep learning, edge computing, artificial intelligence, neural network
Subject Area: TK7885-7895 Computer engineering. Computer hardware |
|---|