The engineering problem
One of three connected research efforts at AUC, alongside ARGUS and EUPB.
Investigate crossing prediction using complementary 2D and 3D information from the LOKI dataset.
Approach & contribution
01
Reconstructed point clouds and extracted pedestrian key-points.
02
Combined 2D image-detection features with 3D key-point features using attention and a linear fusion head.
03
Used a fully connected layer to output crossing probability, with optional YOLO–LiDAR fusion at inference.
Results & boundaries
Implemented the research pipeline; no accuracy or benchmark result is supplied in the current materials.
Publication and evaluation assets remain to be added when available.
Content basis: MAZEN_CONTEXT v1.3, 9 October 2026