UrbanVision is an end-to-end, resource-efficient pipeline for bitemporal 3D change detection designed specifically for edge computing environments. The system compares 3D point cloud scans of the same urban area captured at different times to detect new constructions, demolitions, and structural/environmental changes with high accuracy.
This project was developed and validated on Raspberry Pi 4B, focusing on lightweight models, minimal computation, and practical deployability.
- 🔄 Bitemporal Change Detection using 3D point clouds
- ⚡ Edge-optimized pipeline with low computational overhead
- 🧠 Supports multiple ML/DL models (classical + deep learning)
- 🗂️ Custom parser for non-standard point cloud formats
- 🧪 Model comparison across accuracy, size, and inference speed
- 📦 Ready for real-world urban monitoring & disaster assessment
The pipeline follows these major stages:
- Data Acquisition – Bitemporal 3D scans of the same region
- Data Preprocessing – Cleaning, normalization, and alignment
- Feature Extraction – Geometric and statistical features
- Model Training & Selection – Lightweight and efficient models
- Edge Deployment – Inference on Raspberry Pi 4B
- Dataset Used: Urb3DCD (Simulated urban dataset – Lyon, France)
- Data Type: Bitemporal 3D point clouds (XYZ + RGB)
- Challenge: Non-standard file format (incompatible with Open3D)
- Solution: Custom-built parser to load and preprocess data correctly
UrbanVision experiments with a mix of classical ML and deep learning models, including:
- XGBoost
- Lightweight neural networks
- Optimized classical classifiers
Model selection prioritizes:
- Accuracy
- Inference latency
- Model size (edge constraints)
📌 Pre-trained and optimized models are available under GitHub Releases: 👉 https://github.com/rushi-k12/UrbanVision/releases/tag/models-v1
You can directly download and use these models for inference without retraining.
-
Accurate detection of changed vs unchanged regions
-
Strong performance even with limited compute resources
-
Clear visual distinction:
- 🟥 Red: Changed regions
- 🟩 Green: Unchanged regions
Comparative analysis across models highlights the trade-off between accuracy and efficiency for edge deployment.
- Target Device: Raspberry Pi 4B
- Inference: Fully offline, minimal network dependency
- Deployment Goal: Real-time or near real-time urban monitoring
- 🔧 Model quantization for further size reduction
- ⏱️ Real-time streaming point cloud processing
- 🌍 Validation on real-world LiDAR datasets
- 🤖 Improved generalization across sensors & environments
- Live Demo: https://connect.raspberrypi.com/devices
- YouTube Demo: https://youtu.be/FS80BiFMFN8
This work was presented at:
5th International Conference on Advanced Network Technologies and Intelligent Computing (ANTIC-2025) Paper ID: 396
Authors:
- Rushikesh Kusuma
- Vamshidhar Narsingoji
- Sahil Jaiswal
- Kiran Kumar Pattanaik
If you find this project useful, consider starring ⭐ the repository and exploring the released models.
Feel free to open issues or contribute!