NavigAid-model

Data pipeline and machine-learning models behind NavigAid, an experimental research proejct for a pedestrian navigation app in Boston. The models estimate the probability of a vehicle-pedestrian crash at a location from nearby street-level infrastructure.

Paper: NavigAid: A Route Analysis and Navigation Model for Pedestrian Safety Using Random Forest Algorithm

Approach

• Task: binary classification of whether a sampled point had a pedestrian crash within 10 m. Day and night are modeled separately, and the night model adds streetlights.

• Sampling: 300k points from the OSMnx Boston walking network. Crash and no-crash classes are balanced by undersampling (41,810 day rows, 18,764 night rows).

• Features (counted or measured within 10 m, from Analyze Boston data):

  Feature	Source
  Speed limit (proxy for impact speed)	Street Segments
  Sidewalk width, slope, damaged-area ratio	Sidewalk Inventory
  Crosswalk coverage	Sidewalk Centerline
  Curb ramps	Pedestrian Ramp Inventory
  Street trees	BPRD Trees
  Streetlights (night only)	Streetlight Locations
  Crash labels	Vision Zero

• Model: scikit-learn RandomForestClassifier (300 trees, random_state=42) with isotonic calibration on a held-out calibration split, so predict_proba outputs can be used as crash probabilities. Split: 80/20 train/hold-out, then 80/20 base/calibration within train.

• Other models compared: logistic regression (plain, degree-2, lasso), SVM, and gradient boosting.

Results (random hold-out split, class-balanced)

• Speed limit is the strongest predictor in both models, matching prior crash research. Sidewalk condition, slope, and width follow close behind. Crosswalk coverage also contributes positively to predicted risk, likely because it reflects pedestrian exposure.
• Ramps, lighting, and tree coverage have the smallest effects.

Feature importance

Day	Night

Gini importance from the random forest.

Metrics

Hold-out metrics (random split, so likely optimistic):

	Day	Night
Accuracy	0.965	0.969
ROC AUC	0.991	0.994
Brier score (uncalibrated → calibrated)	0.043 → 0.024	0.044 → 0.028

Limitations

• Metrics are on balanced data, so they may not reflect real-world crash base rates.
• The hold-out is a random split, so nearby points with near-identical features can land in both train and test. Spatial cross-validation is future work.

Other plots (ROC, calibration, confusion matrices, learning curves, partial dependence, etc.) are in results/ and discussed in the paper.

Repo layout

model/
  preprocess/   generate_points.py → filter_boston_crash_data.py → compile_data.py
                (feature_engineering.py, geo_helpers.py are shared helpers)
  models/       random_forest.py plus the comparison models
  utils/        plotting and model helpers
  datasets/     raw city data, sampled points, and the compiled day/night CSVs
final/          trained calibrated models (*.pkl, gitignored)

Reproduce

Run everything from the repo root (scripts use relative paths).

pip install pandas numpy geopandas osmnx shapely scikit-learn shap matplotlib seaborn joblib
python model/preprocess/generate_points.py        # sample Boston walking network
python model/preprocess/filter_boston_crash_data.py
python model/preprocess/compile_data.py           # build day/night CSVs
python model/models/random_forest.py              # train, calibrate, evaluate, save to final/

The compiled CSVs are already in model/datasets/final_csv/, so you can skip straight to training. The paper used Python 3.12.7 and scikit-learn 1.5.2; use the same scikit-learn version when loading saved models, since pickles are version sensitive.

Using the models in the app

Copy the trained models to the app repo as app/backend/models/day_model.pkl and night_model.pkl. The backend averages per-point probabilities along a route to score it.