NeuroPark – 3D Vehicle Detection

Problem

Autonomous parking systems need accurate 3D perception of their surroundings but expensive LiDAR sensors are not always available.

Solution

Developed a pipeline using YOLOv8 for object detection, MiDaS for depth estimation, and point cloud reconstruction to create 3D scene representations from monocular images.

Architecture

The pipeline combines YOLOv8 (object detection), MiDaS (depth estimation), and Open3D (point cloud processing) with PCA/SVD for dimensionality reduction.

Tech Stack

Key Challenges

• Depth estimation accuracy from single images
• Real-time processing requirements
• Handling occlusions and adverse weather