Trajectory Prediction Modeled by Bimodal Bivariate Distribution

Humans move through complicated, crowded situ-ations by abiding social norms, which include respecting others personal space, giving way to traffic, and avoiding collisions. Social connections occur in crowded settings in a variety of ways that are frequently inconspicuous. Deep learning methods have recently outperformed their manually crafted counterparts by understanding human interactions in a broader, data-driven manner. However, the conventional approaches often assume unimodal Gaussian distributions, limiting their ability to capture the multimodal nature of trajectory predictions. A person behind another person is equally likely to follow the person ahead or overtake him, and the future trajectory therefore has two very different future versions. We propose a deep learning-based methodology by proposing a novel approach for trajectory prediction by modeling the distribution of future possibilities as a bimodal bivariate distribution. The model architecture consists of an LSTM architecture with channel-wise attention. The architecture also contains an intent encoding that captures the long-term intent of the person. Apart from maximizing the probability of observing the demonstrated trajectory, the loss function also regularizes the curvature of the predicted trajectory. Further, to counter any skewness in the dataset, we augment the dataset during training by mirroring it on X and Y Axis around the centre of the scene. Experimental results show that the proposed approach outperforms popular approaches and several baselines.