Video Anomaly Detection based on Memory-augmented Adversarial Training Network
Liang Zhang, Shifeng Li, Ye Tian, Xi Luo, Xiaoru Liu · 2024
We tackle the challenge of anomaly detection by employing an adversarial training approach to enhance the autoencoder's reconstruction capabilities, along with a memory module designed to capture patterns of normal data. To refine the recording of details, we integrate entropy loss and a hard shrinkage rectified linear unit (ReLU) into the memory module. Additionally, we incorporate skip connections into our model to prevent the memory module from merely capturing the most prototypical representation patterns, thereby ensuring a comprehensive representation of necessary patterns. We validate our approach using publicly accessible datasets, demonstrating that our method is capable of effectively detecting anomalies from video sequences