BILAE-GAN FRAMEWORK FOR ANOMALY DETECTION IN VIDEO SURVEILLANCE

摘要

© 2023 Little Lion Scientific.In recent years, increasing the use of surveillance cameras with less manpower makes automatic video surveillance systems to become more important. Recent advances in video anomaly identification have mostly focused on improving performance with available datasets. We propose a Bidirectional Long-Short term memory-based Convolutional Autoencoder Generative Adversarial Network (BiLAE-GAN) method for video surveillance. During training the model learns the normal data distribution of data in the Generator and the detection of anomalies in the discriminator. Bidirectional Long-Short term network in Convolutional Autoencoder in Generator for reconstruction, Encoder features of Generated image and real image to discriminator to identify Anomaly. At the anomaly detection phase, anomalies are identified based on reconstruction error and discrimination results. Our proposed method validation benchmark datasets such as UCSD Ped1, UCSD Ped2, and CHUCK Avenue dataset with performance metrics AUC, EER.