Transformers and LSTM for Trajectory Prediction

Predicting the trajectories of pedestrians is a difficult task since it depends on a number of external factors. The context of the scene and the interaction between pedestrians are two of the most important factors. Several methods have been explored in past to tackle this issue transitioning from physics-based models to data-driven models based on sequence based neural networks like RNNs. In this project we have compared the performance of Long-Short terms memory and Transformer Architectures. Since, attention is the most important aspect for trajectory prediction, spatial and temporal context of trajectory are taken into account for making future predictions in interaction aware LSTM and Transformer architecture.We use mean average displacement and final average displacement as comparison metrics for architecture comparison.

Visualization of Predictions from Simple Transformer Network

Visualization of Predictions from Social LSTM Network

Visualization of Predictions from Social Transformer Network