A Data-Driven Framework for Pedestrian Oriented Route Planning Leveraging Deep Learning and Spatial Perception

Official Implementation

A framework for predicting urban perception attributes (beautiful, lively, boring, safe) using machine learning models with delta sensitivity analysis.

Authors

• Pyare Lal Chauhan (LinkedIn)
• Tanishq Kumar Baswal
• Vaibhav Kumar

Affiliation: GeoAI4Cities Lab, IISER Bhopal | Netrica.ai

Pipeline Architecture

┌─────────────────────────────────────────────────────────────────┐
│                    STAGE 1: Feature Extraction                  │
│                       (Optional/Reference)                      │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  Raw Street View Images (~111,268 from Place Pulse 2.0)        │
└─────────────────────────────────────────────────────────────────┘
                              │
                ┌─────────────┴─────────────┐
                ↓                           ↓
┌───────────────────────────┐   ┌───────────────────────────┐
│  Semantic Segmentation    │   │   Object Detection        │
│  (ADE20K - 150 classes)   │   │   (COCO - 80 classes)     │
└───────────────────────────┘   └───────────────────────────┘
                └─────────────┬─────────────┘
                              ↓
         ┌────────────────────┴────────────────────┐
         ↓                                         ↓
┌─────────────────────────┐         ┌─────────────────────────┐
│  Full Extraction        │         │  Finetuned DINOv3       │
│  (150 SS + 80 OD = 230) │         │  (36 features)          │
└─────────────────────────┘         └─────────────────────────┘
                              │
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│                 STAGE 2: Perception Prediction                  │
│                      (This Repository)                          │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ↓
┌─────────────────────────────────────────────────────────────────┐
│  Delta Sensitivity Analysis                                     │
│  7 thresholds (δ = 0.5, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8)          │
└─────────────────────────────────────────────────────────────────┘
                              │
        ┌──────────┬──────────┼──────────┬──────────┐
        ↓          ↓          ↓          ↓
   Random Forest  SVM     XGBoost    RealMLP
                              │
                              ↓
              Perception Predictions & Visualizations
              (112 trained models, 12+ figures)

Pre-extracted features included - start immediately!

Installation

git clone https://github.com/geoai4cities/svi_perception.git
cd svi_perception
source setup_experiment.sh

Quick Start

# Test run (5-15 min)
./run_experiment.sh --test

# Full experiment (2-6 hours)
./run_experiment.sh --full --background

# Monitor progress
./monitor_experiment.sh

Data

Pre-extracted features in Input_Data/dinov3_all_classes/:

• 111,268 images with 36 finetuned DINOv3 features
• 4 perception files: beautiful_input.xlsx, lively_input.xlsx, boring_input.xlsx, safe_input.xlsx

Feature Extraction Options (see Feature_Extraction_Code/):

• Full extraction: 230 features (150 ADE20K segmentation + 80 COCO detection)
• Finetuned DINOv3: 36 optimized features (used in this repo)

Results

Results saved in experiments/<dataset>/<city>/:

• 03_results/metrics/all_results.csv - All 112 experiment results
• 03_results/visualizations/ - Publication-ready figures

Documentation

• QUICK_START.md - Getting started guide
• docs/PIPELINE.md - Complete pipeline details
• docs/FEATURE_EXTRACTION.md - Feature extraction guide

License

Non-commercial / academic use only, under the PolyForm Noncommercial License 1.0.0 — see LICENSE.

Contact

• Email: pyare22@iiserb.ac.in
• Lab: GeoAI4Cities
• Paper: ScienceDirect