NeuroPark - 3D Vehicle Detection
YOLOv8 + MiDaS depth + Open3D pipeline turning single 2D traffic-camera images into oriented 3D vehicle scenes.
A computer vision project led with Neurosoft that explores monocular perception for 3D understanding. The pipeline integrates object detection, depth estimation, point-cloud generation, and spatial alignment to produce usable vehicle scenes without relying on LiDAR as the primary sensor.
Led a 3-person team (with Neurosoft) turning single 2D traffic-camera images into 3D vehicle scenes
Used YOLOv8 detection, MiDaS depth estimation, and Open3D, validated against LiDAR measurements
Reached ~5-6 s per high-resolution image and presented the system at the KPZ25 Engineering Conference
Impact
Validated monocular 2D-to-3D reconstruction, presented at KPZ25
Role
Team Lead / ML Engineer
Timeline
2025
Key tags
Problem
3D scene understanding is useful for autonomous parking and traffic analysis, but richer sensors are expensive and not always available.
Solution
Built a pipeline that uses monocular RGB input, estimates depth, reconstructs point clouds, and localizes vehicles in aligned 3D world coordinates.
Architecture
YOLOv8 handles detection, MiDaS estimates depth, Open3D builds colored point clouds, and PCA/SVD-based geometry constructs oriented 3D boxes aligned with world coordinates.
Challenges
- Depth estimation uncertainty from single-camera input
- Spatial alignment and coordinate calibration across the full pipeline
- Balancing scene quality with processing speed in near-real-time scenarios
Technology stack
Continue exploring
More case studies
Lucid - AI Study App
Helps students turn lecture screenshots and notes into searchable study material, flashcards, and quizzes.
Knowledge Graph RAG (ToPWR)
Helps students and staff query university information spread across pages, departments, and systems.
Medical Chatbot RAG
Provides educational health Q&A from retrieved source content instead of free-form LLM answers.