Transfer learning model for anomalous event recognition in big video data

Roqaia Adel Taha ¹, Aliaa Abdel-Halim Youssif ², Mohamed Mostafa Fouad ²

⁰College of Computing and Information Technology, Arab Academy for Science, Technology and Maritime Transport (AASTMT), Smart Village, Cairo, Egypt. rokaiaadel2020@gmail.com.

Scientific Reports +

|

November 13, 2024

View abstract on PubMed

Summary

This study introduces a semantic key frame extraction algorithm for automated surveillance, significantly reducing video data. The Vision Transformer (ViT_b16) model achieved 95.87% accuracy in recognizing anomalous events.

Area Of Science

Computer Science
Artificial Intelligence
Machine Learning

Background

Video surveillance systems face challenges in accurately recognizing anomalous human activities.
High monitoring costs and operator fatigue limit the effectiveness of traditional CCTV systems.
Automated, real-time event recognition is crucial for enhancing security.

Purpose Of The Study

To develop a semantic key frame extraction algorithm for efficient anomalous event recognition in surveillance videos.
To evaluate the performance of deep learning models (ResNet50, VGG19, EfficientNetB7, ViT_b16) using this novel approach.
To address the issue of large video data volumes by minimizing frame processing requirements.

Main Methods

A semantic key frame extraction algorithm based on action recognition was proposed.
The algorithm was integrated with ResNet50, VGG19, EfficientNetB7, and Vision Transformer (ViT_b16) models.
The models were trained and tested on the UCF-Crime dataset, comprising both normal and abnormal surveillance videos.

Main Results

EfficientNetB7 achieved 86.34% accuracy, VGG19 reached 87.90%, and ResNet50 attained 90.46%.
The Vision Transformer (ViT_b16) model demonstrated superior performance with 95.87% accuracy.
The proposed method significantly improved anomalous event recognition compared to state-of-the-art models.

Conclusions

The semantic key frame extraction algorithm effectively reduces video data while maintaining high accuracy in anomaly detection.
The Vision Transformer (ViT_b16) model shows exceptional promise for real-time anomalous event recognition in surveillance.
This research offers a scalable and accurate solution for automated security systems.