90 % turn-prediction accuracy from smartphone video alone — no training, no maps, no calibration.
Paper (GeoAI 2026) · Dataset (coming soon) · Noah Meißner*, Tim Sieber*, Bernd Ludwig · University of Regensburg
Indoor GPS fails — and IMU-based tracking loses up to 50 % accuracy at turns. We show that horizontal optical flow from a standard smartphone camera predicts turn direction (left / right / straight) at 90 % accuracy across 10 participants on routes up to 399 m, nearly eliminating the turn-accuracy drop from prior work.
For full methodology, dataset details, and results see the paper.
from optical_flow.ZoneFlowPredictor import Predictor
predictor = Predictor(zone_radius=4.5, threshold=1.0, global_bias=0.0472, algrthm="lucas-kanade")
# df must contain columns: x_new, y_new, android_image_filename
direction = predictor.moved(df, zone_pos=(5400, 2100))
# Returns: "links" | "rechts" | "gerade"The full end-to-end pipeline (data loading → sync → prediction → ablation study) is in Zone_Flow_Predictor.ipynb.
A pedestrian turning creates a lateral shift in the camera's field of view — a directional trend in horizontal optical flow. We gate analysis to a zone of radius r around each decision point and classify the zone mean against a threshold t.
| Parameter | Optimal | Description |
|---|---|---|
zone_radius |
4.5 m | Radius around a decision point |
threshold |
1.0 | Min. mean flow to classify a turn |
global_bias |
0.0472 | Camera-drift correction |
Recording setup: HoloLens 2 (6DoF ground truth) + Google Pixel 9 (video + IMU), synchronised via Raspberry Pi 5 at 7.95 ± 0.24 ms latency. Loop closure error: 0.05 % (≈ 8.8 cm) over 176 m.
- 10 participants · 27 trajectories · 3 zone types
- 176 m reference track + 399 m complex track
| Notebook | What it does |
|---|---|
01_Data_Quality.ipynb |
Loop-closure error, outlier detection |
02_Predictor.ipynb |
Per-zone prediction visualisation |
03_Visualisation.ipynb |
Flow signal + trajectory plots |
Zone_Flow_Predictor.ipynb |
Full pipeline: data → LORTO-CV → ablation |
@inproceedings{meissner2026optical,
title = {Seeing Around the Corner: Fusing Visual Flow and Inertial Sensors for Indoor Pedestrian Navigation},
author = {Mei{\ss}ner, Noah and Sieber, Tim and Ludwig, Bernd},
booktitle = {Proceedings of the 1st International Conference on Geospatial Artificial Intelligence (GeoAI 2026)},
year = {2026},
address = {Ghent, Belgium}
}