A CNN-GRU Based Hybrid Approach for Pedestrian Trajectory Prediction

Pedestrian trajectory prediction has emerged as a crucial component in the realm of computer vision and autonomous systems, playing a pivotal role in ensuring safe and efficient autonomous driving. It has diverse applications in surveillance systems, collision avoidance, intersection management and crowd management in public spaces. In autonomous vehicles, predicting pedestrian movements is essential for ensuring safe navigation and decision-making, thereby minimizing the risk of accidents. However, the task of accurately predicting pedestrian trajectories has numerous challenges. Pedestrians exhibit complex and diverse movement patterns influenced by factors such as individual behaviour, environmental conditions and social interactions. Unpredictable events, such as sudden stops, or group dynamics, further compound the difficulty of accurate prediction. In this work, we introduce a CNN-GRU architecture for forecasting pedestrian trajectories. We evaluate this model against several data-driven approaches, each trained independently on the ETH/UCY dataset. Our comparison includes LSTM, Vanilla LSTM, Social LSTM, and GRU model. Results show that the CNN-GRU model achieves superior performance, yielding ADE/FDE scores of 0.46/0.61, outperforming the other models.