Scholarlink

Video anomaly detection (VAD) (Iyengar et al., Artificial intelligence in practice: theory and applications for cyber security and forensics. Springer Nature (Forthcoming)) is essential in intelligent surveillance, helping identify unusual events in video sequences to improve safety and security. This chapter presents a comparative study between a vision language model-based approach utilizing Vision Transformers (ViTs) integrated with the CLIP model and a deep learning (Zhang et al., IEEE Trans Circuits Syst Video Technol 32(8):5427–5437, 2022) approach employing object tracking and Context Conditional Variational Autoencoder (CVAE) for VAD on the NWPU campus dataset. Our vision language model-based approach uses CLIP’s feature association (Wu et al., IEEE Trans Neural Netw Learn Syst 1–14, 2019) capabilities and a custom temporal annotation file for weakly supervised anomaly detection, enhancing detection accuracy (Brooks and Iyengar, Multi sensor fusion: fundamentals and applications with software. Prentice Hall, New Jersey. Number of Pages : 488) through feature alignment between visual and textual embeddings. Conversely, the deep learning approach integrates ByteTrack for tracking and CVAE to incorporate contextual scene information (Vert et al., Introduction to contextual processing – theory and application. CRC Press, pp. 320), improving scene-dependent anomaly detection by distinguishing contextually abnormal events. Both models are evaluated on frame-level AUC and mAP metrics under various IoU thresholds, enabling a robust comparison. Experimental results demonstrate that each approach has distinct advantages depending on the anomaly’s context and complexity, underscoring the importance of architecture choice (Xavier and Iyengar, Introduction to parallel algorithms (Chinese), 263 p. ISBN: 7-11-113390-0; Xavier and Iyengar, Introduction to parallel algorithms. Wiley. Number of Pages: 365; Soloway and Iyengar (eds), Empirical studies of programmers. Ablex, Norwood) in video anomaly detection applications. The deep learning model, which learned patterns without any text labels (unsupervised), outperformed the vision language model, achieving an AUC of 67.5% compared to 60.05%. Even with the additional textual supervision in the vision language model, the deep learning approach proved more effective in identifying anomalies. These findings highlight the potential for unsupervised (Yang et al., Anomaly detection in surveillance videos via memory-augmented frame prediction. In: 2022 international joint conference on neural networks (IJCNN), pp. 1–8, 2022; Wang et al., IEEE Trans Neural Netw Learn Syst 33(6):2301–2312, 2021; Sun et al., Scene-aware context reasoning for unsupervised abnormal event detection in videos. In: Proceedings of the 30th ACM international conference on multimedia, pp 184–192, 2020) learning for this task.

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

Learning Scene Context for Anomaly Detection: A Benchmarking Study of Forward Frame Prediction CNN and Vision Language Model-Based Approach

  • 章节

  • 发表日期：

    2026年

查看全文
  • Preet Kanwal,
  • Shylaja S S,
  • Niroop Karera,
  • Nischal Kashyap,
  • Poorvi Tambakad,
  • Rishab Gongulur,
  • Prasad B Honnavalli

摘要

Video anomaly detection (VAD) (Iyengar et al., Artificial intelligence in practice: theory and applications for cyber security and forensics. Springer Nature (Forthcoming)) is essential in intelligent surveillance, helping identify unusual events in video sequences to improve safety and security. This chapter presents a comparative study between a vision language model-based approach utilizing Vision Transformers (ViTs) integrated with the CLIP model and a deep learning (Zhang et al., IEEE Trans Circuits Syst Video Technol 32(8):5427–5437, 2022) approach employing object tracking and Context Conditional Variational Autoencoder (CVAE) for VAD on the NWPU campus dataset. Our vision language model-based approach uses CLIP’s feature association (Wu et al., IEEE Trans Neural Netw Learn Syst 1–14, 2019) capabilities and a custom temporal annotation file for weakly supervised anomaly detection, enhancing detection accuracy (Brooks and Iyengar, Multi sensor fusion: fundamentals and applications with software. Prentice Hall, New Jersey. Number of Pages : 488) through feature alignment between visual and textual embeddings. Conversely, the deep learning approach integrates ByteTrack for tracking and CVAE to incorporate contextual scene information (Vert et al., Introduction to contextual processing – theory and application. CRC Press, pp. 320), improving scene-dependent anomaly detection by distinguishing contextually abnormal events. Both models are evaluated on frame-level AUC and mAP metrics under various IoU thresholds, enabling a robust comparison. Experimental results demonstrate that each approach has distinct advantages depending on the anomaly’s context and complexity, underscoring the importance of architecture choice (Xavier and Iyengar, Introduction to parallel algorithms (Chinese), 263 p. ISBN: 7-11-113390-0; Xavier and Iyengar, Introduction to parallel algorithms. Wiley. Number of Pages: 365; Soloway and Iyengar (eds), Empirical studies of programmers. Ablex, Norwood) in video anomaly detection applications. The deep learning model, which learned patterns without any text labels (unsupervised), outperformed the vision language model, achieving an AUC of 67.5% compared to 60.05%. Even with the additional textual supervision in the vision language model, the deep learning approach proved more effective in identifying anomalies. These findings highlight the potential for unsupervised (Yang et al., Anomaly detection in surveillance videos via memory-augmented frame prediction. In: 2022 international joint conference on neural networks (IJCNN), pp. 1–8, 2022; Wang et al., IEEE Trans Neural Netw Learn Syst 33(6):2301–2312, 2021; Sun et al., Scene-aware context reasoning for unsupervised abnormal event detection in videos. In: Proceedings of the 30th ACM international conference on multimedia, pp 184–192, 2020) learning for this task.