Future Frame Prediction Using a U-Net-Based Wasserstein Generative Adversarial Network (WGAN) Approach

This paper presents a novel approach to activity prediction using Generative Adversarial Networks (GANs). Activity prediction plays a crucial role in applications such as surveillance and human-computer interaction. Previous works in activity prediction have struggled with long-term predictions and have primarily relied on skeleton data, which lack rich visual context and fine-grained motion details. To address these challenges, we propose a GAN-based model that predicts future frames by learning complex motion dynamics, overcoming the limitations of relying solely on skeleton data. The proposed model is evaluated on two benchmark datasets, UT-Interaction and UCF11, using two evaluation metrics: Mean Squared Error (MSE) and Structural Similarity Index (SSIM). We also analyze the model’s performance with different loss functions, demonstrating its efficacy. The results indicate significant improvements in prediction accuracy compared to existing methods. Future work will explore alternative algorithms to better capture temporal dependencies and consider additional datasets to improve prediction accuracy and robustness in real-world applications.