Background

Camera traps are essential for wildlife monitoring, but transmitting raw images is energy-intensive. Running object detection locally on embedded hardware enables intelligent filtering and event-based alerts. This project investigates deploying a vision model for animal detection on a microcontroller with an integrated camera interface.

Key Topics

Embedded computer vision Model compression and optimization Image preprocessing on constrained hardware Firmware development in C/C++ Performance and energy evaluation

Learning Outcomes

Understand challenges of vision AI on microcontrollers. Gain experience with embedded vision pipelines and toolchains. Implement and evaluate real-time inference on-device. Benchmark accuracy, speed, and power.

Project Outline

Literature Review: Explore TinyML approaches for wildlife monitoring. Hardware Familiarisation: Learn to use a modern microcontroller board with integrated camera and dedicated AI acceleration. Model Selection: Choose a small animal detection model or train a simple classifier. Optimisation: Adapt the model for embedded deployment using compression and quantisation. Firmware Development: Implement camera capture, preprocessing, and inference. Testing & Benchmarking: Evaluate detection accuracy, latency, and energy use. Report: Summarize findings and propose improvements.

Possible Research Questions

What are the limitations of real-time object detection on microcontrollers with constrained memory? How does input resolution and model size affect detection accuracy and inference speed? Can embedded vision systems reliably detect animals in varying lighting and environmental conditions?

Requirements

Interest in embedded systems and AI. Motivation to learn C/C++ and microcontroller development. Basic Python knowledge for model preparation.