Improving Deep Unsupervised Anomaly Detection by Exploiting VAE Latent Space Distribution