Loading...
Search
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
Digital Object List
Bookmark
Review List
814 viewed
تشخیص شرایط غیرعادی در ویدیوهای دوربین های مدار بسته
اعتمادی نایین، محمد علی Etemadi Naeen, Mohammad Ali
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 58874 (05)
- University: Sharif University of Technology
- Department: Electrical Engineering
- Advisor(s): Bagheri Shouraki, Saeed; Mohammadzadeh, Hoda
- Abstract:
- With the rapid growth of surveillance systems and the increasing demand for effective analysis of large-scale video data, human-centric anomaly detection has become a critical problem in computer vision. However, this task faces several significant challenges, including the high variability of human behaviors, the complexity and noise of real-world scenes, class imbalance, models’ reliance on irrelevant background context, and the scarcity of labeled data for rare anomalous events. Many existing approaches fail to deliver stable and generalizable performance in real-world operational settings due to excessive dependence on contextual cues or inadequate modeling of temporal dynamics. To address these challenges, this research proposes an effective framework for detecting human-related anomalies in surveillance videos, based on the integration of human-centric preprocessing and deep spatio-temporal modeling. The core idea is to deliberately focus on behaviorally relevant regions while suppressing distracting scene information through a human-centered processing pipeline. In this framework, YOLO-World is employed for open-vocabulary person detection, ByteTrack is used for identity-aware tracking, and a background-blurring preprocessing stage is applied to ensure that the learning process concentrates primarily on human behavior. Spatial features are extracted using an ImageNet-pretrained InceptionV3 network and subsequently mod- eled over time through a BiLSTM architecture to capture temporal dependencies and motion dynamics. The proposed framework was evaluated on a five-class subset of the UCF-Crime dataset. Experimental results demonstrate that the model achieves a mean test accuracy of 92.41% across three independent runs, outperforming prior methods. Per-class F1-scores are all above 0.85, while macro (0.90) and weighted (0.92) averages confirm robustness under class imbalance. Stable convergence behavior and the absence of significant overfitting further indicate the effectiveness of the proposed strategy in improving generalization capability and model stability
- Keywords:
- Anomaly Detection ; Surveillance Cameras ; Deep Learning ; Machine Vision ; Artificial Intelligence ; YOLO-World Algorithm
-
محتواي کتاب
- view
- مقدمه
- کارهای پیشین
- مقدمه
- تعریف ناهنجاری در ویدئوهای نظارتی
- روشهای مبتنی بر ویژگیهای دستی
- روشهای مبتنی بر یادگیری عمیق
- روشهای مبتنی بر یادگیری نظارتشده
- روشهای بدوننظارت و خودنظارتی
- مدلهای مکانی-زمانی در تحلیل ویدئو
- یادگیری چندوجهی و همترازی دیداری-زبانی
- Dataset تشخیص ناهنجاری ویدئویی
- پژوهشهای مرتبط تشخیص ناهنجاری در ویدئوهای نظارتی
- جایگاه پژوهش حاضر و شکافهای تحقیقاتی
- معرفی الگوریتم ارائه شده
- نتایج
- نتیجهگیری
- مراجع
- واژه نامه انگلیسی به فارسی
- واژه نامه فارسی به انگلیسی