Social GAN: Socially Acceptable Trajectories with Generative Adversarial Networks This paper presents a Generative Adversarial Network (GAN) approach for predicting socially acceptable pedestrian trajectories in crowded scenes. The method combines sequence prediction and GANs to model human motion behavior, which is inherently multimodal and socially constrained. The key contributions include a novel pooling mechanism that aggregates information across people and a variety loss that encourages diverse predictions. The model outperforms prior work in terms of accuracy, variety, collision avoidance, and computational complexity on several real-world datasets. The approach uses a GAN-based encoder-decoder framework with a pooling module that captures social interactions. The generator predicts future trajectories based on past motion data, while the discriminator evaluates the socially acceptable nature of these predictions. A variety loss is introduced to encourage diverse outputs, and a novel pooling mechanism is used to model global interactions between people. The model is evaluated on two public datasets (ETH and UCY) and shows state-of-the-art performance in terms of accuracy and speed. It outperforms baselines such as linear regression, LSTM, and S-LSTM in trajectory prediction tasks. The model is able to generate diverse socially acceptable trajectories, demonstrating its ability to handle the multimodal and socially constrained nature of human motion in crowded environments. The method is also shown to be significantly faster than S-LSTM, with a 16x speed improvement. The model's ability to generate diverse trajectories is validated through qualitative analysis of collision avoidance scenarios, where it successfully predicts socially acceptable paths. The latent space exploration shows that certain directions in the latent space are associated with direction and speed, indicating the model's ability to generate diverse samples. The method is designed to be data-driven and learns social norms through the pooling mechanism, enabling it to produce globally consistent and socially compliant trajectories.Social GAN: Socially Acceptable Trajectories with Generative Adversarial Networks This paper presents a Generative Adversarial Network (GAN) approach for predicting socially acceptable pedestrian trajectories in crowded scenes. The method combines sequence prediction and GANs to model human motion behavior, which is inherently multimodal and socially constrained. The key contributions include a novel pooling mechanism that aggregates information across people and a variety loss that encourages diverse predictions. The model outperforms prior work in terms of accuracy, variety, collision avoidance, and computational complexity on several real-world datasets. The approach uses a GAN-based encoder-decoder framework with a pooling module that captures social interactions. The generator predicts future trajectories based on past motion data, while the discriminator evaluates the socially acceptable nature of these predictions. A variety loss is introduced to encourage diverse outputs, and a novel pooling mechanism is used to model global interactions between people. The model is evaluated on two public datasets (ETH and UCY) and shows state-of-the-art performance in terms of accuracy and speed. It outperforms baselines such as linear regression, LSTM, and S-LSTM in trajectory prediction tasks. The model is able to generate diverse socially acceptable trajectories, demonstrating its ability to handle the multimodal and socially constrained nature of human motion in crowded environments. The method is also shown to be significantly faster than S-LSTM, with a 16x speed improvement. The model's ability to generate diverse trajectories is validated through qualitative analysis of collision avoidance scenarios, where it successfully predicts socially acceptable paths. The latent space exploration shows that certain directions in the latent space are associated with direction and speed, indicating the model's ability to generate diverse samples. The method is designed to be data-driven and learns social norms through the pooling mechanism, enabling it to produce globally consistent and socially compliant trajectories.