Autonomous-3D-Mapping

Overview

This project presents the development, modeling, and experimental validation of an autonomous end to end indoor 3D reconstruction system using a monocular camera- equipped differential-drive mobile robot. The goal is to control the platform through the environment on a predefined path to capture images of the scene, then implement the 3D reconstruction pipeline to build a virtual 3D Model of the space. A comprehensive dynamic model of the robot—encompassing mechanical and electrical subsystems—was identified using input-output experiments and refined via genetic algorithm optimization. The work incorporates the usage of FIR filter to remove the noise from feedback signals (such as Encoders). The control architecture employs a cascaded scheme with a Lyapunov-based outer loop and LQR-based inner loop. To optimize performance metrics like path accuracy and energy efficiency, a Predictive Neural Network (PNN) was trained to map target positions to optimal control gains (𝐾𝑝 , 𝐾𝜃), enabling real-time adaptive control. For the 3D reconstruction part, the hierarchical localization pipeline is implemented to build a 3D model of the environment form a set of captured images. Empirical benchmarking identified DISK (feature extractor) and LightGlue (matcher) as the optimal vision components. These are fed into a reconstruction pipeline using COLMAP and OpenMVS, producing high-quality 3D models of indoor scenes. The system is implemented on a Raspberry Pi–Arduino platform running ROS and validated both in simulation and on a real robot. The robot autonomously navigates pre-defined discretized paths and captures structured image arcs at predefined waypoints. This end-to-end framework bridges control theory with practical computer vision, demonstrating how low-cost mobile robots can autonomously map environments with high spatial accuracy using only monocular vision and optimized motion planning.

TL;DR: End-to-end autonomous 3D reconstruction of indoor environment using DDWMR

✨ Key Features

• Integrating neural network-optimized hybrid control system
• Deploying the optimal feature extractor and matcher—DISK and LightGlue—for dense and accurate 3D point cloud generation.
• Identifying the robot system using "System-Identification" optimized through Genetic Algorithm
• Filtering the Feedback signals with customized filter (FIR) for reliable localization and control.

Demo

This video shows the end-to-end autonomous 3D reconstruction Process

3D_reconstruction.mp4

Methodology

1. Setting the Discretized Path:

Since the control algorithm is designed for a stabilization task—e.i., guiding the robot from its initial position to a target location—the first step in our project should be to define a discretized path through the relevant environment.

2. Navigating through the environment:

This step focuses on navigating the robot through the environment to capture images of surrounding objects. The process is carried out using the control scheme described in the following repo "https://github.com/B-A-IntelliBots/AI-optimized-LQR-Lyapunov-Control-System".

3. 3D Reconstrution:

This step involves constructing a 3D virtual model of the indoor environment using images captured by the robotic platform. The reconstruction process employs the DISK Feature Extractor and LIGHTGLUE Feature Matcher to achieve a highly accurate and densely detailed final output. The choice of this feature extractor–matcher combination is based on extensive comparative analysis supported by statistical metrics, as detailed in the following repository "https://github.com/B-A-IntelliBots/3D-Reconstruction".

---

🚀 Quick Start

Prerequisites

• OS/Tooling: Python ≥3.10 / MATLAB R2014a / ROS1 noetic / CATIA V5R21
• System deps: cmake

Installation

• Clone the repository:

  $ git clone https://github.com/B-A-IntelliBots/Autonomous-3D-Mapping.git 3d-mapping

---

📊 Results & Benchmarks

the performance of the proposed System State-Space Identification method—using MATLAB’s System Identification Toolbox in combination with a Genetic Algorithm (GA)—is validated according to the following criteria:

🔹 1. Loss / Error Metrics

Quantitative performance is evaluated using:

• Euclidean Loss (custom-defined residual error distance)

Model Type	Euclidean Loss
GA-Optimized	0.1189

🔹 2. Computational Performance

• GA converged in ~[782] generations
• Execution time: ~[47 Sec] on Intel i5 CPU

---

📁Project Structure

├── src/ # Core codes with the robot CATIA-designed components
├── data/ # Sample datasets
├── results/ # Plots and evaluation metrics
├── README.md # This file

---

📖 Citation

@software{myproject2025,
  author    = {Ali Deeb, Bisher Alsaleh}, Supervised by: {Prof. Iyad Hatem}
  title     = {Autonomous 3D Mapping},
  year      = {2025},
  publisher = {GitHub},
  url       = {https://github.com/B-A-IntelliBots/Autonomous-3D-Mapping}
}