AnomLite: Efficient binary and multiclass video anomaly detection

Anomaly detection in video surveillance is critical for ensuring public safety, as manual monitoring of numerous video feeds is often challenging and prone to human error. Security operators can struggle to maintain focus over prolonged periods, leading to missed events or delayed responses. An automated system can help mitigate this by providing continuous, reliable monitoring and rapid anomaly detection. In response to these challenges, we propose AnomLite, a lightweight yet effective model designed to detect anomalies in video streams through a hybrid architecture. The initial layers of MobileNetV2 are employed for efficient spatial feature extraction, capturing low-to-mid-level features such as edges and textures, while the last hidden state of an LSTM processes temporal dependencies to identify patterns indicative of anomalies. AnomLite is competitive due to its computational efficiency, requiring only 11 million parameters, and its robustness, achieving a ROC AUC of 0.99, Average Precision of 0.99 and F1-Score (Weighted) of 0.92 and outperforming comparable models in anomaly detection tasks. These attributes make AnomLite suitable for real-time applications in resource-constrained environments like urban centers and public transport hubs. The code could be found here: https://github.com/AnnaZverev/UCF_Crime.git • Light hybrid model for video anomaly detection. • Achieved 8% improvement over state-of-the-art methods on UCF-Crime dataset. • Addressed both binary and multiclass classification with appropriate metrics.