Skip to content

CS 206: Computer Vision & Embedded AI
Course Code CS 206
Course Name Computer Vision & Embedded AI
Department Computer Science
Semester Offered Odd (Term 3 - Shanghai)
Tuition Hours 30 hours (Theory) + 10 hours (Lab)
Course Level Intermediate to Advanced
Pre-requisite CS 105: Practical Machine Learning
Co-requisite -
Course Objective Most AI systems that interact with the real world need to see. Vision is not just another feature. It is often the primary interface between machines and reality.

This course introduces students to computer vision from a practical, systems-first perspective. Instead of going deep into theory, the focus is on understanding just enough of neural networks to use them effectively, and then building real vision systems that run on hardware.

Students will learn how to process images and video, detect and track objects, and deploy models on edge devices such as Raspberry Pi and Jetson Nano. They will also explore hybrid systems where edge devices communicate with the cloud for compute-heavy tasks.

By the end of the course, students will be able to build real-time vision-enabled systems, directly contributing to their Term 3 goal of building AI-powered hardware products.
Course Philosophy This course emphasizes
  • Use before theory
  • Systems over models
  • Real-time performance over offline accuracy
Students are not expected to become deep learning researchers. They are expected to become builders who can use vision as a capability inside a working system.
Course Learning Outcomes Upon successful completion of this course, students will be able to:
  • Understand the intuition behind CNNs and how they process images.
  • Use pre-trained vision models for tasks like classification and detection.
  • Implement object detection, segmentation, and tracking systems.
  • Work with camera inputs and video streams in real time.
  • Deploy vision models on edge devices such as Raspberry Pi and Jetson Nano.
  • Optimize models for latency and performance in constrained environments.
  • Build hybrid edge-cloud systems for vision-based applications.
  • Integrate vision into hardware products, such as smart wearables and robotic systems.
Course Author Sagar Udasi
MSc Statistics and Data Science with Computational Finance from The University of Edinburgh.
Contact: sagar.l.udasi@gmail.com
Course Organiser TBD
Details will be updated before course commencement.
No. Lecture Title Concepts Covered Lecture Objective
01 How Do Machines See The World? Images as data, pixels, channels Build intuition about visual data representation.
02 Why Neural Networks Work On Images CNN intuition, filters, feature extraction Introduce CNNs without heavy math.
03 From Pixels To Predictions Image classification basics Build first vision model quickly.
04 Seeing More Than Labels Object detection concepts Move from classification to localization.
05 Drawing Boxes Around The World Bounding boxes, detection pipelines Enable practical detection systems.
06 Understanding Every Pixel Segmentation basics Teach fine-grained visual understanding.
07 Following Things That Move Object tracking Build dynamic vision systems.
08 Working With Real Cameras Camera interfaces, video streams Connect models to real-world inputs.
09 Real-Time Vision Is Hard Latency, frame rates, optimization Handle constraints in live systems.
10 Pre-trained Models Save Time Transfer learning in vision Accelerate development using existing models.
11 Running Models On Tiny Devices Edge AI basics, constraints Introduce deployment on embedded systems.
12 Raspberry Pi And Jetson In Action Device setup, inference pipelines Hands-on deployment on hardware.
13 When Edge Is Not Enough Edge-cloud architecture Combine local and remote compute.
14 Building Video Processing Pipelines Streaming, buffering, processing Structure real-time data pipelines.
15 Optimizing For Speed Quantization, pruning basics Improve performance on limited hardware.
16 Integrating Sensors Beyond Cameras Multi-sensor systems Expand beyond vision into robotics context.
17 From Vision Model To Product Feature System integration Connect vision outputs to product logic.
18 Case Study: Vision In A Smart Device Real-world system breakdown Link directly to Term 3 wearable projects.
19 Building Your Own Vision System End-to-end system design Students build their own working system.
20 Demo Day: Does Your System See In Real Time? Live demos, evaluation Validate real-time performance and usability.

Lab Sessions (7 Sessions)

No. Lab Title Concepts Covered Objective
L1 Camera Setup That Actually Works Camera interfacing Enable students to capture real-world data.
L2 Your First Vision Model Image classification Build and run a basic vision model.
L3 Detect Objects In The Wild Object detection Apply detection to real scenes.
L4 Track Movement In Real Time Object tracking Build dynamic systems.
L5 Deploy On Edge Device Raspberry Pi / Jetson setup Run models on hardware.
L6 Edge Meets Cloud Hybrid systems Build edge-cloud pipelines.
L7 Build Your Vision Product End-to-end system Create a working vision-based hardware feature.
Component Weightage
Vision Assignments (4 total) 25%
Lab Performance & Builds 25%
Final Project: Vision-Enabled Hardware System 30%
Live Demo + System Evaluation 20%
Type Resource Provider
Lecture CS231n: Convolutional Neural Networks for Visual Recognition Stanford
Lecture Practical Deep Learning for Coders fast.ai
Reading Deep Learning for Computer Vision Adrian Rosebrock
Documentation OpenCV Documentation opencv.org
Documentation PyTorch Vision pytorch.org
Practice Kaggle Computer Vision Competitions kaggle.com