Real-Time AI Object Detection Model & Demo

1. Problem Statement & Requirements

Academic research papers in Artificial Intelligence (AI) and Machine Learning (ML) can often be flagged by examiners as "purely theoretical" if they are only presented in static Jupyter Notebooks comparing raw accuracy metrics. Thesis panels expect students to present a live, **interactive Demo application** with a clean user interface.

• Functional Requirements: Process image inputs from webcam streams or file uploads, classify objects across customized labels, and render bounding box outputs with corresponding confidence scores.
• Non-functional Requirements: Guarantee smooth processing throughput to ensure latency-free real-time execution on standard hardware configurations.

2. Solution & System Architecture

We suggested utilizing a pre-trained **MobileNet-SSD** or **YOLOv8-nano** model, then applying **Transfer Learning** on the student's customized image dataset to optimize training time and data labeling efforts:

We implemented the **Non-Maximum Suppression (NMS)** algorithm to filter out duplicate bounding box overlap regions, guaranteeing clean visual tracking outputs during the defense panel presentation.

3. Technologies Used

4. Results & Demo

The final system achieves stable processing rates and smooth execution on standard consumer laptops, classifying objects accurately with a high confidence score in the sandbox environment:

• Minimal processing latency, running smoothly on integrated CPU/GPU architectures.
• Clean script codebase documented with detailed explanations for easy presentation.
• Supplied detailed model training documentation for prospective re-evaluation.