Trajectory-Forecasting

Abstract

Trajectory prediction specifically involves predicting the future position, velocity, direction, and other state information of a target object, given its historical or current motion trajectory. This discipline holds significant importance across various domains, finding extensive applications in robotics, autonomous driving, unmanned aerial vehicles (UAVs), motion analysis, and more. In the context of the challenge at hand, we are tasked with performing pedestrian trajectory prediction on the JRDB dataset, with the objective of generating predictions for pedestrians' future trajectories over a duration of 4.8 seconds at a frequency of 2.5Hz. Historically, conventional approaches have largely relied on individual trajectory prediction models, often failing to harness the complementary performance of multiple models.To address this limitation, we propose a multi-model fusion-based trajectory prediction approach. This approach amalgamates various trajectory prediction algorithms, including NSP-SFM and Y-Net, which have demonstrated strong performance on publicly available datasets. Additionally, we leverage object detection data to supplement information about pedestrian counts, resulting in a substantial reduction of the EFE Card metric to 1.293. Our approach yielded promising results in the ICCV23 Trajectory Prediction Competition.