This paper proposes a novel approach for video anomaly detection based on future frame prediction. The method leverages the difference between predicted future frames and their ground truth to identify abnormal events. Unlike existing methods that focus on minimizing reconstruction errors of normal events, this approach introduces a temporal constraint by enforcing optical flow consistency between predicted and ground truth frames. This temporal constraint is the first to be incorporated into video prediction tasks for anomaly detection. The method also incorporates appearance constraints (intensity and gradient loss) and motion constraints (optical flow loss) to improve prediction quality. Additionally, a Generative Adversarial Network (GAN) is integrated to enhance the realism of predicted frames. The proposed framework is evaluated on multiple datasets, including a toy dataset and publicly available datasets, demonstrating its effectiveness in detecting anomalies. The method outperforms existing approaches in terms of robustness to uncertainty in normal events and sensitivity to abnormal events. The framework is implemented using a U-Net architecture for frame prediction, with a combination of appearance, motion, and adversarial training constraints. The results show that the method achieves higher accuracy in anomaly detection compared to existing methods, particularly in distinguishing between normal and abnormal events. The method is also tested on a toy dataset, where it successfully detects abnormal events even in the presence of uncertainty in normal events. The framework is efficient, with an average running time of 25 fps, and is implemented using NVIDIA GPUs and TensorFlow. The results demonstrate the effectiveness of the proposed method for video anomaly detection.This paper proposes a novel approach for video anomaly detection based on future frame prediction. The method leverages the difference between predicted future frames and their ground truth to identify abnormal events. Unlike existing methods that focus on minimizing reconstruction errors of normal events, this approach introduces a temporal constraint by enforcing optical flow consistency between predicted and ground truth frames. This temporal constraint is the first to be incorporated into video prediction tasks for anomaly detection. The method also incorporates appearance constraints (intensity and gradient loss) and motion constraints (optical flow loss) to improve prediction quality. Additionally, a Generative Adversarial Network (GAN) is integrated to enhance the realism of predicted frames. The proposed framework is evaluated on multiple datasets, including a toy dataset and publicly available datasets, demonstrating its effectiveness in detecting anomalies. The method outperforms existing approaches in terms of robustness to uncertainty in normal events and sensitivity to abnormal events. The framework is implemented using a U-Net architecture for frame prediction, with a combination of appearance, motion, and adversarial training constraints. The results show that the method achieves higher accuracy in anomaly detection compared to existing methods, particularly in distinguishing between normal and abnormal events. The method is also tested on a toy dataset, where it successfully detects abnormal events even in the presence of uncertainty in normal events. The framework is efficient, with an average running time of 25 fps, and is implemented using NVIDIA GPUs and TensorFlow. The results demonstrate the effectiveness of the proposed method for video anomaly detection.