Xingnan Zhou

Mapping 70,541 Waymo scenarios onto real-world OpenStreetMap intersections to reveal what intersection types the training data actually covers and where geographic blind spots exist. Uses 48-dimensional star pattern fingerprints (90% top-1 accuracy) with temporal continuity validation — chaining overlapping 9.1s windows into continuous drives (483 connections, 56% OSM path concordance vs 3.9% baseline) to enable corridor-level traffic analysis.

A waterflow lane graph extraction method with cross-attention fusion that achieves +26.7% minADE and +42.7% miss rate improvement on 89K Waymo intersection scenarios. Architecture-agnostic across LSTM and Transformer backbones.

A Turn-Aware LSTM that encodes cumulative heading changes and one-hot maneuver indicators for intersection trajectory prediction. Reduces FDE by 15–20% for turning maneuvers vs. vanilla LSTM, with real-time inference (~2.5 ms) on UAV-captured intersection data in Montreal.

A spatial attention visualization framework for Transformer-based trajectory prediction. Discovers that cyclists receive up to 73% less attention than vehicles — a safety blind spot with 88.1% miss rate. Introduces counterfactual attention analysis on the Waymo dataset (MTR-Lite, 8.48M params).

A symmetric dual-camera LiDAR fusion framework combining PointPillar and CenterPoint with drone and forward-camera YOLOv8 detections. Achieves +4.4% mAP@0.5 improvement (sign test p = 0.001) on a CARLA Town10HD dataset with 10-seed validation.

Ranked #15 on the Waymo E2E Driving Challenge — ADE 1.28m (3s) / 2.99m (5s), RFS 7.70, within 0.11m of #1 (NTR). Trained on a single RTX 4090 with QLoRA, no reinforcement learning. GRPO RL on Google TPU Research Cloud coming next.