A preprint posted to arXiv on 25 September 2026 describes a pedestrian trajectory prediction model that cuts errors exceeding a 1-metre safety tolerance by roughly one-third at a single urban intersection, using time-to-collision (TTC) and crossing-zone context.
Standard metrics such as average displacement error (ADE) and final displacement error (FDE) can hide the occasional large mistake that would matter most for a collision-avoidance system. The authors argue that counting exceedances—predictions whose error surpasses a defined tolerance—is a more safety-relevant yardstick.
The work draws on the inD (Intersection Drone) dataset, which records naturalistic trajectories at one intersection. Trained with a 1.6-second observation window to forecast 2.4 seconds ahead, several neural architectures were compared. A position-only long short-term memory (LSTM) served as the baseline, while a pooled LSTM separately encoded TTC histories and crossing-zone indicators before merging them with pedestrian positions.
The study also introduces a weighted loss that penalises large coordinate-wise errors more heavily. Applied to the position-only LSTM, this weighted loss reduced ADE from 0.210 m to 0.190 m and FDE from 0.550 m to 0.503 m. The number of predictions exceeding the 1 m threshold fell 34.8% for ADE and 19.8% for FDE.
The final model, which adds TTC and crossing-zone information on top of the weighted loss, achieved ADE of 0.184 m and FDE of 0.491 m. Compared with the safety-oriented position-only LSTM, it delivered further reductions in exceedance counts of 33.5% (ADE) and 6.3% (FDE).
The preprint has not been peer reviewed, and its evaluation is confined to one intersection from one dataset. The authors note that “broader validation across pedestrians, sites, and datasets is required.” No independent testing or real-world deployment data accompany the results.