Enhancing Surveillance Video Abnormal Behavior Detection Using Deep Convolutional Generative Adversarial Network With GRU Model

Automatic detection of unusual behavior in videos is a challenging task. This challenge comes from its complexity and the wide range of applications it covers. Several deep learning approaches have been proposed to address this challenge. This includes recent generative methods that use deep convolutional generative adversarial networks (DCGAN). The DCGAN model has gained high research attention recently due to its performs well in extracting spatial features and solve class imbalance issue to detect abnormalities. However, a DCGAN is unstable during training and has low performance owing to its inability to capture the long-term temporal dependency between sequences of video frames. In this study, we propose a novel gated recurrent unit (GRU)-based DCGAN architecture to improve the training stability and performance of a DCGAN model for abnormal video behavior detection. The proposed model was trained using UCSD Ped1, UCSD Ped2, CUHK Avenue, and ShanghaiTech benchmark anomaly dataset. Compared to the DCGAN model, the proposed GRU-based DCGAN model improved the detection accuracy and area under the curve (AUC) by an average of 19.91% and 8.57%, respectively. Compared with the 3D-DCGAN model, the GRU-based DCGAN model improved the detection accuracy and AUC by an average of 7.67% and 3.73%, respectively. Furthermore, the GRU-based DCGAN model stabilized from epoch 10 and converged at epoch 38, whereas the other models remained unstable and did not converge at epoch 50. The findings highlight that the combination of GRU to enhance temporal modeling within a DCGAN framework is a logical extension to improve training stability and performance for abnormal video behavior detection.