Deep Learning for Anomaly Detection: A CNN-LSTM Autoencoder Approach

Due to the significant growth of video surveillance installations in recent years, video anomaly detection has attracted considerable interest in the security sector. Detecting anomalies in video is a crucial task within the realm of smart security. The intricacy and diversity of real-world data can present challenges for conventional methods, leading to decreased accuracy. In this paper, we introduce a robust system for identifying anomalies in surveillance footage by employing a deep-learning-driven unsupervised anomaly detection technique that relies on a reconstruction approach. Our architecture employs an autoencoder with CNN and LSTM layers to extract local and motion characteristics from the video frames. The evaluation of the proposed model is conducted using benchmark datasets, UCSD Pedestrian 1 and UCSD Pedestrian 2. The performance of our model was noticeably superior to that of the other competing models, with AUC and EER values of 99.1% and 6.4% for Ped1, 98.5% and 4.4% for Ped2, respectively.