Atten-GAN: Pedestrian Trajectory Prediction with GAN Based on Attention Mechanism

Predicting pedestrian trajectories in dynamic scenarios is extremely challenging due to the mobility and flexibility of pedestrian motion. However, most existing methods cannot fully extract the interaction information between pedestrians. In this paper, a generative adversarial network model-based attention mechanism (Atten-GAN) was proposed to model social relationships of the interaction information between pedestrians. The Atten-GAN is composed of a generator and a discriminator. The generator predicts multiple possible future trajectories according to the past trajectories of pedestrians. The discriminator scores the trajectories according to the trajectories input, determines whether the trajectories are ground-truth or generated by the generator, and then facilitates the generator to generate trajectories in line with social norms. Atten-GAN introduces an attention pooling module to allocate the influence weight of pedestrians in the scene, which can fully extract pedestrian interaction information. In addition, to solve the problem associated with how the GAN network gradient is easy to disappear and difficult to train, the noise decreasing with time is introduced into the loss function of the discriminator during the training. The comparison experiments on ETH and UCY datasets showed that Atten-GAN could not only provide a variety of socially acceptable prediction trajectories in accordance with the social norms but was also was superior to the existing generative model-based methods in the prediction accuracy. The Atten-GAN model had a significant improvement in prediction accuracy and improved the training effects.

Automated summary

This paper proposes a method based on generative adversarial network model and attention mechanism to predict pedestrian trajectories in dynamic scenarios. The method can extract the interaction information between pedestrians and allocate the influence weight of pedestrians using an attention pooling module. Experimental results demonstrate the superiority of this method in prediction accuracy compared to existing methods.