Real-Time Multi-Camera Vehicle Tracking System

Real-time multi-camera vehicle tracking integrating object detection, multi-view fusion, and custom tracking logic — based on a BMW Group manufacturing case study.

Built for a BMW Group Plant Hams Hall manufacturing case study, this system fuses feeds from multiple cameras to maintain consistent vehicle identities across views in real time. It includes a full end-to-end ML pipeline — data collection, annotation, and YOLO model training — along with homography-based spatial mapping and a multi-threaded PySide6 GUI running at 25–30 FPS.

Tech Stack

PythonYOLOOpenCVNumPyPySide6

Features

Real-time multi-camera object detection and tracking at ~25–30 FPS
End-to-end ML pipeline: data collection, annotation, and YOLO model training
Consistent object identity maintained across multiple camera views
Homography-based spatial mapping for cross-camera analysis
Multi-threaded architecture with interactive PySide6 visualisation

Challenges

Maintaining consistent object identities across overlapping and non-overlapping camera views.
Achieving real-time throughput within the constraints of a multi-threaded Python application.
Collecting and annotating a custom dataset representative of the industrial environment.
Applying homography transforms accurately under varying lighting and perspective conditions.

Learnings

Deep understanding of multi-view fusion and object re-identification strategies.
Practical experience building and tuning YOLO models on custom industrial datasets.
Proficiency in designing multi-threaded Python pipelines for real-time video processing.
Hands-on application of homography and projective geometry for spatial analysis.

Links

Live Github