NAVIENTA is a driving safety support prototype that classifies risk into three levels:
LOW (safe), MEDIUM (be careful), and HIGH (danger).

This prototype uses Random Forest and a real-time dashboard simulation to show risk changes over time.
The current focus is safe distance risk as a first development stage.

Why Random Forest (instead of rule-based logic)?

• Data-driven: learns patterns from data, not only hand-made rules
• Flexible: easier to extend when adding new input features
• Easier to evaluate: performance can be measured with clear metrics, and feature importance helps explain model behavior

Baseline reference

• Uses the two-second rule as an initial guideline for safe following distance (NY State DMV Driver’s Manual, p. 47)

Next scope (planned)

• Integrate LiDAR TF350 data
• Expand to overtaking risk by considering distance to the front vehicle and the oncoming vehicle

NAVIENTA adalah prototipe sistem pendukung keselamatan berkendara yang membagi risiko menjadi 3 level:
LOW (aman), MEDIUM (hati-hati), dan HIGH (bahaya).

Prototipe ini menggunakan Random Forest dan dashboard real-time (simulasi) untuk menampilkan perubahan risiko dari waktu ke waktu.
Fokus saat ini adalah risiko jarak aman sebagai tahap pengembangan awal.

Mengapa Random Forest (dibanding sistem berbasis aturan)?

• Berbasis data: belajar pola dari data, tidak hanya aturan manual
• Fleksibel: lebih mudah dikembangkan saat menambah fitur input
• Mudah dievaluasi: performa dapat diukur dengan metrik yang jelas, dan feature importance membantu menjelaskan perilaku model

Acuan awal

• Menggunakan two-second rule sebagai pedoman awal jarak aman (NY State DMV Driver’s Manual, hlm. 47)

Rencana pengembangan

• Integrasi data LiDAR TF350
• Perluasan ke risiko menyalip dengan mempertimbangkan jarak kendaraan depan dan kendaraan lawan arah

The dataset used in this project is publicly available on Kaggle and can be accessed via the following link: [https://www.kaggle.com/datasets/nigelwilliams/ngsim-vehicle-trajectory-data-us-101]

Please note that the dataset is not included in this repository in accordance with Kaggle’s data usage policy.

Dataset yang digunakan dalam proyek ini tersedia secara publik di Kaggle dan dapat diakses melalui tautan berikut: [https://www.kaggle.com/datasets/nigelwilliams/ngsim-vehicle-trajectory-data-us-101]

Dataset tidak disertakan langsung dalam repository ini sesuai dengan kebijakan penggunaan data Kaggle.

• src/prepare_data.py — data loading, cleaning, unit conversion, feature engineering, baseline labels
• src/train_model.py — Random Forest training, metrics, confusion matrix, model saving
• src/make_insights.py — demo CSV generation, smoothed risk timeline, high-risk segments, feature importance
• notebooks/train_navienta.ipynb — notebook version (EDA + results)
• app_realtime.py — Streamlit simulated real-time dashboard

Note: data/, outputs/, and model/ are not tracked by Git (see .gitignore).

Raw columns used (proxy dataset):

• Vehicle_ID, Global_Time → timestamp
• Spacing → d_front_m (meters)
• Vehicle_Velocity → v_ego_ms (m/s)
• Vehicle_Acceleration → acc_ms2 (m/s²)
• Headway → headway_s (seconds) for baseline labeling
• (optional) Lane_ID, Frame_ID

Derived features (feature engineering):

• dt
• closing_speed (distance closing rate)
• ~1-second rolling features (≈10 rows at 10Hz):
  min_d_front_1s, med_d_front_1s, max_close_1s, time_to_reach_1s

Prototype labels:

• risk_class ∈ {LOW, MEDIUM, HIGH} created from headway rule (baseline)

Important: headway_s is not used as a model feature to avoid leakage.

Kolom mentah yang digunakan (dataset proxy):

• Vehicle_ID, Global_Time → timestamp
• Spacing → d_front_m (meter)
• Vehicle_Velocity → v_ego_ms (m/s)
• Vehicle_Acceleration → acc_ms2 (m/s²)
• Headway → headway_s (detik) untuk baseline labeling
• (opsional) Lane_ID, Frame_ID

Fitur turunan (feature engineering):

• dt
• closing_speed (laju mendekat / jarak mengecil)
• fitur rolling ~1 detik (≈10 baris pada 10Hz):
  min_d_front_1s, med_d_front_1s, max_close_1s, time_to_reach_1s

Label prototipe:

• risk_class ∈ {LOW, MEDIUM, HIGH} dibentuk dari aturan headway (baseline)

Penting: headway_s tidak dipakai sebagai fitur model untuk menghindari leakage.

Install dependencies:

pip install -r requirements.txt