Scholarlink

This study presents a robust deep learning-based framework for criminal activity detection in surveillance video footage, utilizing advanced spatiotemporal modeling techniques. The proposed model integrates state-of-the-art architectures, including ResNet, DenseNet, and EfficientNet for spatial feature extraction, with Long Short-Term Memory (LSTM) networks to capture temporal dynamics. Keyframe extraction using FFmpeg and ffprobe significantly reduces data complexity while retaining critical information, enabling efficient processing of large-scale video datasets. Hyperparameter optimization via Optuna further enhances model performance. The framework was evaluated on the UCF-Crime dataset, which includes 13 real-world anomaly classes, achieving an accuracy of 84%. This marks a significant improvement in precision and computational efficiency compared to existing approaches. Key contributions include addressing challenges related to large data volumes, limited computational resources, and the complexity of classifying diverse criminal activities. The findings emphasize the importance of temporal dynamics in anomaly detection, paving the way for further advancements. This research provides a scalable and high-performance solution to enhance public safety and support law enforcement operations.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient Criminal Activity Detection Using Spatiotemporal Deep Learning: A Novel Approach Leveraging UCF-Crime Dataset and Advanced Video Processing Techniques

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ahmed Mohamed Elessawy,
  • Ahmed Salem,
  • Soha Safwat

摘要

This study presents a robust deep learning-based framework for criminal activity detection in surveillance video footage, utilizing advanced spatiotemporal modeling techniques. The proposed model integrates state-of-the-art architectures, including ResNet, DenseNet, and EfficientNet for spatial feature extraction, with Long Short-Term Memory (LSTM) networks to capture temporal dynamics. Keyframe extraction using FFmpeg and ffprobe significantly reduces data complexity while retaining critical information, enabling efficient processing of large-scale video datasets. Hyperparameter optimization via Optuna further enhances model performance. The framework was evaluated on the UCF-Crime dataset, which includes 13 real-world anomaly classes, achieving an accuracy of 84%. This marks a significant improvement in precision and computational efficiency compared to existing approaches. Key contributions include addressing challenges related to large data volumes, limited computational resources, and the complexity of classifying diverse criminal activities. The findings emphasize the importance of temporal dynamics in anomaly detection, paving the way for further advancements. This research provides a scalable and high-performance solution to enhance public safety and support law enforcement operations.