3D Point Cloud Perception Pipeline

This project demonstrates how to process, segment, and transform raw LiDAR data into meaningful 3D scene understanding for autonomous driving tasks.

---

Overview

The goal of this project is to create a 3D Machine Learning pipeline capable of:

• Detecting lane lines using reflectance data.
• Segmenting the road plane and isolating obstacles.
• Performing 3D object detection via clustering and bounding box fitting.

I wanted to bridge theory with practice: understand how algorithms like RANSAC and DBSCAN actually work on point clouds, and combine them into one working perception pipeline.

---

What I Built

1. Lane Line Detection

I experimented with LiDAR reflectance values to estimate lane markings using two simple but powerful techniques:

1. Thresholding – Filtering high reflectance points to visualize lane lines.
2. Region of Interest (ROI) – Keeping only reflectance values within the lane region for a cleaner lane estimation.

This step highlights how even basic geometric filtering can reveal meaningful road features.

---

2. 3D Object Detection Pipeline

For 3D object detection, I implemented a multi-step pipeline using the Open3D library:

1. Voxel Downsampling: Reduce point density for faster computation.

2. Ground Plane Segmentation (RANSAC): Separates the dominant plane (the road) from the rest of the scene.

3. Clustering (DBSCAN): Group object points based on spatial proximity.

4. 3D Bounding Boxes: Draw 3D boxes around each detected object (clusters).

This pipeline turns raw, noisy data into a structured 3D understanding of the environment.

Results

• Lane lines detected using reflectance thresholding.
• Road plane isolated via RANSAC segmentation.
• Multiple 3D objects detected and visualized with bounding boxes.