Lightweight Real-Time Anomaly Detection and Prediction for Surveillance Videos Using a Self-Supervised Hybrid Autoencoder and Dynamic Graph CNN-LSTM Framework

Abstract

Anomaly detection in surveillance videos is a critical task for ensuring public safety, requiring both high accuracy and real-time efficiency. This study presents a novel lightweight anomaly detection framework that integrates a convolutional Autoencoder for feature extraction and a Graph CNN-LSTM for spatiotemporal anomaly classification. The proposed model is trained and evaluated on the large-scale UCF-Crime dataset, achieving state-of-the-art results with a multiclass classification accuracy of 98%, a multiclass ROC-AUC of 0.99, and a binary anomaly detection ROC-AUC score of 0.91. To enhance computational efficiency, the model incorporates structured pruning and quantization-aware training, reducing memory and inference overhead without compromising performance. Additionally, a self-supervised learning (SSL) prediction module is integrated, contributing to anomaly detection with a very low computational cost (~0.14 ms per frame). The results demonstrate that our optimized model significantly outperforms prior methods in both classification and efficiency. In addition, to validate its suitability for real-time applications, the model underwent extensive time and computational complexity analysis where the inference time per frame for the final optimized model was measured at just ~1.22 ms, with a stable constant-time complexity. These results confirm its practicality for online real-world surveillance environments while boosting robust detection performance and minimal computational costs.