Edge-Deployable Multi-Class Surveillance Video Anomaly Classification Using a MobileNetV2–LSTM Spatial–Temporal Framework

Abstract

Automated surveillance systems require models that can recognize abnormal events accurately while maintaining efficient inference on resource-constrained hardware. This study proposes a lightweight spatial–temporal classification framework that integrates MobileNetV2 and Long Short-Term Memory networks for identifying five surveillance activity classes: Normal, Fighting, Burglary, Road Accident, and Explosion. Surveillance videos were preprocessed through frame sampling, resizing, normalization, and fixed-length sequence construction. MobileNetV2 was employed to extract compact frame-level spatial features, while the LSTM modeled temporal dependencies across consecutive frames. The framework was evaluated using accuracy, class-wise precision, recall, F1-score, macro F1-score, weighted F1-score, confusion-matrix analysis, and probability-based anomaly scoring. Experimental results on 463 video sequences show that the proposed model achieved an accuracy of 97.41%, a macro F1-score of 0.9612, and a weighted F1-score of 0.9744. Most classification errors occurred between visually similar human-centered anomaly classes, particularly Fighting and Burglary, while Normal and Explosion sequences were recognized with very high reliability. The trained model was converted into TensorFlow Lite format and deployed on a Raspberry Pi 4, achieving approximately 6–8 frames per second with an average latency of 120–150 ms per frame. These findings indicate that the proposed MobileNetV2–LSTM framework provides an effective balance between multi-class event recognition, temporal representation learning, and near-real-time edge inference. However, further validation on larger and more diverse surveillance datasets is required to confirm its generalizability under unconstrained operational conditions.