Project Description Template

Table of Content

Project Timeline: Jan 2026 ~ May 2026

Final Result / Demo Visual

Introduction

Project Context

Blank Image for Project Overview

This project was developed during a software/AI robotics bootcamp in South Korea as a team of five. Our team chose to build Neighbot, a software prototype for a neighbourhood patrol robot that could use AI vision to detect safety-related situations.

The goal was to build a vision-to-operator system in which camera input could be processed by an AI model, converted into actionable detection information, and presented to an operator via a monitoring platform. Although we did not build a physical robot, the project simulated the software system that could support a mobile patrol robot in detecting objects such as weapons, cigarettes, or fallen-person situations.

Project Goal

My Role

My role / contribution

I served as the project manager, overseeing project planning, system design, and team coordination throughout development.

Challenge 1: Structuring the Development Path for a Vision-to-Operator System

Starting From a Broad Concept Without a Clear System Path

The project began with a broad concept: use AI vision to help a patrol robot detect safety-related situations and report useful information to an operator. However, before implementation, the team needed to clarify how the robot, AI model, server, GUI, database, and operator workflow would integrate into a single system.

Challenge / Problem Visual 1

Turning the Concept Into a System Scenario

To make the idea buildable, I wrote the system scenario before development went too far. This defined how visual input should move from camera detection to AI processing, server handling, GUI display, database storage, and operator review.

Challenge / Problem Visual 2

Breaking the Workflow Into Module Responsibilities

Once the scenario was defined, I helped translate the system flow into module responsibilities and team tasks. Using Jira, Confluence, meetings, sprint planning, and action items, I helped clarify who was responsible for each part and what needed to be completed each week.

Challenge / Problem Visual 3

Keeping Development Aligned With Integration Goals

Because each module was developed separately, I scheduled integration testing around specific workflow targets rather than treating each part as complete in isolation. When issues appeared across the AI model, server, GUI, database, or interfaces, I helped the team stay on track by prioritizing the most immediate system-level problem.

Challenge / Problem Visual 3

Creating a Clear Path From Concept to Working Prototype

This structure helped the team move from an open-ended AI-robot idea toward a working vision-to-operator prototype. As a result, the team was able to finalize a system that could process detection results, present them to the operator, and support review through the GUI and database workflow.

Challenge / Problem Visual 3

Challenge 2: Improving AI Detection Through Better Training Data

Discovering That Public Training Data Was Not Enough

We first tried training the model with public datasets from Roboflow-type sources, but the results were unreliable. The model failed to detect the actual objects we had and also produced false detections, such as recognizing arms or a long ceiling light as a knife, a phone as a gun, and a pencil as a cigarette.

Challenge / Problem Visual 1

Identifying Data Quality as the Main Detection Problem

After researching the issue, we realized the problem was not only the model itself but the quality and relevance of the training data. The model could only detect objects reliably when the training images closely matched the object’s shape, angle, colour, lighting, distance, and background.

Challenge / Problem Visual 2

Creating Custom Data for the Objects We Actually Used

Because the public datasets did not match our demonstration objects well enough, we decided to create our own training data. I filmed scenario-specific actions with the target objects so the model could learn from the actual knife, gun-like object, cigarette, and emergency-related poses we needed to demonstrate in real time.

Challenge / Problem Visual 3

Filtering and Balancing the Dataset for More Reliable Detection

I also helped review and filter a large number of images (around 10,000!) to remove low-quality or irrelevant data. While preparing the dataset, we considered variation and balance across object classes so the model would not become overly biased toward one target or fail under slightly different viewing conditions.

Challenge / Problem Visual 3

Proving the Concept Within a Controlled Prototype Scope

The final model was not a universal detector, but it became reliable enough to detect the specific objects and situations we trained it on. This proved the core concept: with relevant, balanced, scenario-specific training data, the AI vision system could detect selected safety-related events and provide operators with timely alerts and visual cues to support monitoring and response.

Challenge / Problem Visual 3

Conclusion

Outcome

By the end of the project, our team completed a working AI vision-monitoring prototype that could detect selected safety-related objects and situations, present detection information via an operator-facing GUI, and store detection records for later review. The system was not a universal real-world detector, but it successfully demonstrated the concept within the controlled scope of a prototype.

Outcome Vissual

Key Lesson

The most valuable lesson from this project was that AI model performance depends heavily on the quality, relevance, and balance of the training data. I learned that working with a deep learning model often feels like working with a black-box system: even when the model output is wrong, the cause is not always obvious, so improvement requires careful data review, repeated testing, and strategic trial and error.

Extra Visual