Identification of Abnormal Activities in Surveillance Scenes using Deep Auto Encoder and Sequential Extreme Learning Machine

Video anomaly detection is an essential task in computer vision, focused on spotting unexpected or unusual events in surveillance footage. This study presents a hybrid framework that combines a Deep Autoencoder (AE) with a Sequential Extreme Learning Machine (ELM) to enhance anomaly detection. The AE plays a crucial role in capturing local features, executing nonlinear transformations, and compressing time-series data, which significantly minimizes the need for manual feature design and labeling. Furthermore, the Sequential ELM offers quick training and adaptability, allowing for dynamic analysis of real-time data. The framework employs convolutional neural network (CNN)-based autoencoders to analyze video data, where convolutional and pooling layers enhance feature extraction, ensuring effective anomaly classification. By utilizing a bottleneck architecture, the AE generates compact representations that are subsequently used to train the Sequential ELM. Experiments conducted on benchmark datasets, such as UCSD Peds2, Avenue, and ShanghaiTech, highlight the effectiveness of the proposed method compared to state of art models. The findings reveal extensive performance in detecting both spatial and temporal abnormal events while handling challenges presented by dynamic environments and diverse data. Various evaluation metrics, including Area Under the Curve (AUC) and Peak Signal-to-Noise Ratio (PSNR), indicate that the model achieved more results compared to existing methods in terms of accuracy and resource efficiency. The proposed system demonstrates high precision in identifying anomalous activities, offering a scalable solution for real-time automated surveillance. Future work will focus on integrating more advanced deep learning architectures and broadening evaluations across various datasets.