Bachelor's Thesis · 2019
Autonomous Unmanned Vehicle System
AUVS
End-to-end traffic sign recognition pipeline with real-time GPIO-based hardware control for autonomous vehicle guidance. Integrates computer vision for live sign detection with embedded control logic that translates visual input into vehicle motion without human intervention.
Abstract
AUVS addresses the challenge of bridging high-level perception with low-level vehicle control in a resource-constrained embedded environment. The system takes a live camera feed as its sole input and produces physical vehicle movement as output, with all decision-making happening on-device in real time.
The core contribution is the perception-to-actuation pipeline: a trained sign classifier feeds directly into a GPIO control layer, which drives the motors governing vehicle direction and speed. This removes the need for a human operator in the command loop while keeping the hardware footprint small enough for deployment on a physical prototype.
The system was validated against Malaysian road sign variants across stop, proceed, and turning commands, demonstrating reliable closed-loop behaviour in a controlled test environment.
System Components
Vision Pipeline
Live camera feed processed through a computer vision model for real-time traffic sign detection and classification
Sign Recognition
Trained classifier distinguishing stop, go, turn, and hazard signs under varied lighting and angle conditions
Control Layer
GPIO interface translating classified visual inputs into discrete motor control signals for vehicle actuation
Embedded Runtime
Lightweight inference loop running on constrained hardware with deterministic latency requirements
Key Contributions
- Designed a closed-loop perception-action pipeline connecting camera input to hardware actuation with no human-in-the-loop
- Built a custom training dataset of traffic signs captured under Malaysian road conditions
- Implemented real-time sign detection with sub-second response latency on embedded hardware
- Validated system reliability across stop, proceed, left-turn, and right-turn command classes
Technology
Python
Language
OpenCV
Computer Vision
Raspberry Pi
Hardware
GPIO
Embedded Control
TensorFlow Lite
Inference
NumPy
Data Processing
Research Context
This thesis was completed as the capstone project for the Bachelor of Computer Science (Hons.) programme in 2019. It was an early exploration of autonomous systems before the widespread availability of large pretrained vision models, requiring the pipeline to be built with more constrained tooling.
The work established a foundation in embedded AI and hardware-software integration that later fed into work on AI deployment, production inference systems, and real-time decision pipelines in enterprise contexts.
Bachelor's Thesis · Computer Science (Hons.) · 2019
Background