M.Tech Transportation Systems Engineering · IIT Guwahati
Working on physics-informed deep learning for trajectory prediction in heterogeneous, lane-free traffic.
Finishing my M.Tech thesis on HetSpA — a class-conditioned, bicycle-decoded attention model that predicts multimodal vehicle motion on lane-free Indian roads, with every trajectory physically drivable by construction. Pooled best-of-six minADE 0.436 m over 12,762 vehicles under a leakage-free vehicle-ID protocol, plus a diagnostic that asks what in the scene actually carries the predictive signal. Manuscript in preparation for Transportation Research Part C. → trajectory-prediction-indian-traffic
Also in the project: a protocol-honesty study on NGSIM — the same model loses ~18% of its apparent accuracy when vehicle-identity leakage is removed from the customary split. Evaluation protocols matter more than leaderboards.
Motion forecasting · Heterogeneous & lane-free traffic · Physics-informed deep learning · Multimodal prediction · Graph attention · Honest evaluation protocols · Autonomous vehicles