Parallel Symmetric Appearance-Motion Framework With Diffusion and Refinement Blocks for Video Anomaly Detection System

Abstract

Video anomaly detection is crucial in network security, application performance monitoring, and quality control. It recognizes unexpected patterns or behaviors in video footage, allowing for threat detection, app optimization, and product quality improvement. Many deep learning models effectively detect video anomaly detection but have some limitations, such as more time computation and model complexity. To address these issues, this paper proposes the Parallel Symmetric Appearance-Motion framework with Diffusion and Refinement blocks (PSAM-DRB) for detecting video abnormalities. The proposed model's initial step is to pre-process input videos to accentuate anomalous activities through video frame selection. Spatial and temporal Residual Inception-based autoencoder extracts multi-level features and optical flow maps in video frames. Feature decoding is performed using motion- and appearance-dominated branches. A Diffusion Strengthening and Intermodal Refinement block enhances feature representation through cross-scale augmentation and cross-modality interaction. Finally, a fusion module combines the upper and lower branches to detect video anomalies. In this evaluation, the proposed model using the UCF-Crime dataset achieved an accuracy of 99.19%. Finally, the proposed PSAM-DRB framework provides a robust and efficient method for identifying anomalies in video data, with applications in a variety of industries.