In the evolving field of video surveillance, this study introduces Dynamic Crowd Surveillance (DCS)-Detect, an advanced deep learning model designed for real-time detection of suspicious activities across diverse scenarios. Leveraging a 3D ResNet-18 convolutional neural network (CNN), DCS-Detect effectively captures both spatial and temporal patterns in video frames, enabling accurate identification of anomalous behaviors. Customized to classify 13 distinct anomaly types from the DCSASS dataset, DCS-Detect addresses the challenges posed by dynamic environments by integrating a comprehensive data preprocessing pipeline. This pipeline includes frame sampling, augmentation, normalization, and optical flow analysis, all of which enhance the models generalization capabilities. Rigorous experimentation demonstrates DCS-Detects high performance, achieving an accuracy, precision, recall, and F1 score of 9 8. 6 5%, underscoring its robustness and reliability.
This work was supported partially by the BK21 FOUR program of the National Research Foundation of Korea funded by the Ministry of Education (NRF5199991514504).
