🅿️ ParkSense AI

AI-Driven Parking Intelligence for Parking-Induced Congestion

Bengaluru Traffic Police · Flipkart GRiD

Detect illegal-parking hotspots · Quantify their impact on traffic flow · Target enforcement

---

🚦 The Problem

On-street illegal & spillover parking near commercial areas, metro stations and events chokes carriageways and intersections. Today enforcement is patrol-based and reactive, there is no heatmap of violations vs. congestion impact, and it is hard to prioritise enforcement zones.

ParkSense AI turns 298,450 anonymised parking-violation records into an AI decision system that detects hotspots, quantifies their traffic-flow impact, predicts where they will recur, and outputs a deployable patrol plan.

---

🎯 Key Results

Metric	Value	Meaning
🎯 Congestion concentration	62.7%	Top 20 of 263 hotspots hold 62.7% of citywide impact
🔮 Risk-model R²	0.496	Explains ~50% of variance in expected parking load
📈 Beats baseline (MAE)	+28.3%	0.200 vs 0.278 — better than "use the cell average"
🚓 Operational capture	70%	Patrol the top-10% predicted cells → catch 70% of violations
🗺️ Hotspot coverage	96.1%	Share of all impact that sits inside detected hotspots

Validated on a time-based hold-out — trained on the first 80% of dates, scored on the last 20% it never saw.

---

🧠 What Makes It Different

A typical submission	🅿️ ParkSense AI
Heatmap of violation counts	Congestion Impact heatmap (flow-weighted)
Descriptive — what happened	Predictive — what will happen
"Here are the hotspots"	"Send Unit-03 to KR Market, 18:00–21:00"
Black-box accuracy claim	Future-validated, auditable, +28.3% over baseline

---

🖼️ Dashboard Walkthrough

1 · Command Centre & Congestion-Impact Heatmap

Live KPIs, a severity-weighted heatmap, and the ranked hotspot table. The sidebar filters re-scope every metric in real time (city → any police station).

2 · Congestion Impact Score

The CIS leaderboard plus a per-hotspot radar breakdown of the five score components — fully auditable for enforcement officers.

3 · When & Where (Temporal Intelligence)

Hour × day-of-week heatmap, monthly trend and the offending vehicle / violation mix — answering when to deploy patrols.

4 · Predictive Risk Model

Accuracy metrics, feature importance and a what-if predictor of tomorrow's hotspots for any day + time-window.

5 · Enforcement Plan

A deployable patrol schedule — each unit gets a zone, time window and day, with a one-click CSV export.

---

🏗️ Architecture

flowchart TD
    A["Raw CSV · 298k violations"] --> B["data_pipeline.py<br/>clean · IST tz · features"]
    B --> P[("violations_clean.parquet")]
    P --> C["hotspot_detection.py<br/>weighted DBSCAN (haversine)"]
    C --> D["congestion_impact.py<br/>Congestion Impact Score"]
    P --> E["forecasting.py<br/>hour × dow demand profiles"]
    P --> F["risk_model.py<br/>LightGBM · 25 features"]
    D --> G["enforcement.py<br/>patrol optimiser"]
    E --> G
    F --> G
    D --> H{{"Streamlit Command Centre"}}
    E --> H
    F --> H
    G --> H

Loading

---

🔮 The Predictive Risk Model

We predict the expected severity-weighted parking load of every ~278 m cell for each (day-of-week × time-window) — a stable, recurring demand surface a planner can actually act on.

flowchart LR
    subgraph FEATS["25 spatio-temporal features"]
      direction TB
      P["⏱️ Period<br/>cyclical time"]
      Cl["📍 Closeness<br/>cell history"]
      Tr["📈 Trend<br/>drift"]
      Sp["🗺️ Spatial<br/>neighbour + Getis-Ord Gi*"]
      Cx["🚗 Context<br/>severity · vehicle mix"]
    end
    FEATS --> M["LightGBM<br/>Tweedie objective"]
    M --> O["Expected load per<br/>cell × day × window"]
    O --> R["Top-10% cells =<br/>70% of violations"]

Loading

The feature design follows ST-ResNet's closeness + period + trend + external decomposition, plus spatial-autocorrelation (neighbour / Getis-Ord Gi*) features — see Research Foundation.

---

📐 Congestion Impact Score (the core idea)

Instead of counting tickets, each violation is weighted by how much it chokes traffic flow, then blended into one auditable 0–100 score per hotspot.

flowchart LR
    A["Volume · 0.30"] --> Z{{"Congestion Impact Score<br/>0 – 100"}}
    B["Severity · 0.25"] --> Z
    C["Junction · 0.20"] --> Z
    D["Persistence · 0.15"] --> Z
    E["Peak · 0.10"] --> Z
    Z --> T["Tiers:<br/>Critical · High · Medium · Low"]

Loading

CIS = 100 × ( 0.30·Volume + 0.25·Severity + 0.20·Junction
            + 0.15·Persistence + 0.10·Peak )

Severity is a domain weight per violation type, e.g. PARKING NEAR TRAFFIC LIGHT/ZEBRA = 1.0, PARKING IN A MAIN ROAD = 0.95, DOUBLE PARKING = 0.90 … PARKING ON FOOTPATH = 0.40.

---

📚 Research Foundation

Technique	Source	Why we used it
Spatio-temporal decomposition (closeness + period + trend + external)	ST-ResNet — Zhang, Zheng & Qi, AAAI 2017 (arXiv:1610.00081)	Blueprint for the feature groups
Spatial autocorrelation (Getis-Ord Gi*, Moran's I)	Getis & Ord 1992; Anselin 1995; Tobler's First Law 1970	Neighbour / spatial-lag features
Density-based clustering (DBSCAN)	Ester, Kriegel, Sander & Xu, KDD 1996	Organic hotspots, no preset count
Gradient boosting + Tweedie loss	LightGBM — Ke et al., NeurIPS 2017	Zero-inflated, non-negative load target

---

⚙️ Tech Stack

Python 3 · pandas / numpy · scikit-learn (DBSCAN) · LightGBM · SciPy · Folium + streamlit-folium · Plotly / matplotlib · Streamlit · pyarrow

Engineering highlights: live KPIs recomputed from data each render · st.fragment + returned_objects=[] so map zoom never reruns the app · ThreadPoolExecutor parallel artefact loading · cached loaders · one-command launcher.

---

🚀 Quickstart

pip install -r requirements.txt

# one command: build artefacts if missing, then open the dashboard
python run.py

Run stages explicitly

python run.py build       # (re)build data, hotspots, CIS, model, plan, maps
python run.py dashboard   # launch the Streamlit command centre
# or:  python src/build_all.py  &&  streamlit run app/dashboard.py

---

🗂️ Project Structure

ParkSense-AI/
├─ run.py                     # one-command launcher (build + dashboard)
├─ src/
│  ├─ config.py               # paths, congestion weights, CIS weights, grids
│  ├─ data_pipeline.py        # clean · IST tz · parse violations · features
│  ├─ hotspot_detection.py    # weighted DBSCAN (haversine)
│  ├─ congestion_impact.py    # Congestion Impact Score + tiers
│  ├─ forecasting.py          # hour×dow demand profiles · patrol windows
│  ├─ risk_model.py           # LightGBM spatio-temporal risk (25 features)
│  ├─ enforcement.py          # rank + optimise patrol allocation
│  ├─ visualize.py            # Folium maps + figures
│  └─ build_all.py            # runs the whole pipeline
├─ app/dashboard.py           # Streamlit command centre (5 tabs)
├─ .streamlit/config.toml     # theme + server settings
├─ img/                       # dashboard screenshots (this README)
├─ outputs/                   # generated artefacts (data, model, maps, kpis)
├─ requirements.txt
└─ README.md

---

🔐 Data & Ethics

Data: jan to may police violation_anonymized*.csv — anonymised records, Bengaluru (Nov 2023 – Apr 2024). Only anonymised IDs are used — no PII. The model targets places and times, not people or vehicles.

Built to turn reactive patrols into targeted, data-driven enforcement. 🚦