.png){kind=icon}
.png){kind=icon}
Return
Synergizing deep learning for smart video surveillance: cutting-edge approaches to abnormal behavior detection
.png){kind=icon}
PRE
AI
N
H
DOI：10.1080/02533839.2026.2642700.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The rapid expansion of data across multiple domains has heightened the demand for robust Video Anomaly Detection (VAD) systems that support reasoning-based surveillance. Such systems are critical for ensuring public safety and preventing crime by detecting irregular activities in video streams. However, traditional approaches often struggle to handle complex multimedia environments and the wide diversity of anomaly types. To address these challenges, this paper proposes a hybrid deep learning model tailored for intelligent surveillance. The framework employs Transfer Learning with a pre-trained ResNet model to extract spatial information from video frames efficiently, thereby reducing computational complexity and minimizing dependence on large-scale labeled datasets. To capture temporal dynamics, Long Short-Term Memory (LSTM) networks are integrated, enabling the identification of anomalies that evolve across sequences of frames. Furthermore, Generative Adversarial Networks (GANs) are incorporated to enhance robustness by modeling the distribution of normal behaviors and applying reconstruction error as a measure of abnormality. The proposed approach is evaluated on widely used benchmark datasets, including UCSD Ped1, UCF Crime, and ShanghaiTech Experimental results demonstrate that the system effectively detects diverse anomalies, such as crowd irregularities, traffic violations, abandoned objects, fighting, and theft, while achieving high accuracy with a low false alarm rate.
Keywords:
Shanghaitech
video anomaly detection (VAD)
resnet
transfer learning (TL)
generative adversarial networks (GAN)
long short term memory (LSTM)
University of Central Florida (UCF) crime
Journal
J
IF:
1.2
Papers:
122
Citations:
1.1K
Organization
Cited Papers