Anomaly Detection in Video Surveillance for Unusual Behavior Identification
Intekhab Ahmad, Sandhya Rani Sahoo, Ratnakar Dash · 2024
“Anomaly detection in video surveillance” identifies unusual behavior in camera footage, boosting security. State-of-the-art methods use both normal and abnormal videos to train, reducing manual monitoring but potentially causing false alarms. This paper highlights a learning algorithm that trains on both normal and anomalous videos to learn to detect anomalies. This paper employs 3D ResNet-101 for feature extraction, facilitating the capture of rich spatiotemporal information from the videos. Subsequently, it uses the MIL framework with weakly labeled training data. Within this approach, individual video segments are viewed as instances, and regular and anomalous videos are viewed as bags. In order to automatically create a deep anomaly ranking model, we finally include multiple instance learning (MIL). This model assigns high anomaly scores to video segments exhibiting unusual behavior, enabling the effective identification of potential security threats.