The paper introduces a novel approach to anomaly detection in videos by leveraging future frame prediction. Unlike existing methods that primarily focus on minimizing reconstruction errors of training data, this work proposes a framework that predicts future frames and compares them with ground truth to identify abnormal events. The key contributions include: 1. **Future Frame Prediction Framework**: The method predicts future frames using a U-Net architecture, incorporating spatial and temporal constraints. 2. **Spatial and Temporal Constraints**: The framework enforces both appearance (intensity and gradient) and motion (optical flow) constraints to ensure high-quality predictions for normal events. 3. **Generative Adversarial Network (GAN)**: A GAN module is integrated to generate more realistic future frames, enhancing the overall performance. 4. **Evaluation**: Extensive experiments on toy datasets and public datasets (CUHK Avenue, UCSD Pedestrian, ShanghaiTech) validate the effectiveness of the method, showing superior performance over existing methods in terms of robustness and sensitivity to anomalies. The paper demonstrates that the proposed method is more suitable for anomaly detection as it aligns with the concept that normal events are predictable, while abnormal events do not conform to expected patterns. The method's ability to handle uncertainties in normal events and its sensitivity to abnormal events are highlighted through various experiments and comparisons with state-of-the-art methods.The paper introduces a novel approach to anomaly detection in videos by leveraging future frame prediction. Unlike existing methods that primarily focus on minimizing reconstruction errors of training data, this work proposes a framework that predicts future frames and compares them with ground truth to identify abnormal events. The key contributions include: 1. **Future Frame Prediction Framework**: The method predicts future frames using a U-Net architecture, incorporating spatial and temporal constraints. 2. **Spatial and Temporal Constraints**: The framework enforces both appearance (intensity and gradient) and motion (optical flow) constraints to ensure high-quality predictions for normal events. 3. **Generative Adversarial Network (GAN)**: A GAN module is integrated to generate more realistic future frames, enhancing the overall performance. 4. **Evaluation**: Extensive experiments on toy datasets and public datasets (CUHK Avenue, UCSD Pedestrian, ShanghaiTech) validate the effectiveness of the method, showing superior performance over existing methods in terms of robustness and sensitivity to anomalies. The paper demonstrates that the proposed method is more suitable for anomaly detection as it aligns with the concept that normal events are predictable, while abnormal events do not conform to expected patterns. The method's ability to handle uncertainties in normal events and its sensitivity to abnormal events are highlighted through various experiments and comparisons with state-of-the-art methods.