This project focuses on detecting and classifying wall surface texture in real time to support automated inspection in manufacturing-like environments. The objective was to build a perception system capable of distinguishing smooth and rough wall surfaces under varying lighting and deployment constraints.
I fine-tuned a YOLOv5-based object detection model on a custom dataset of 1,300 wall surface images, carefully curated to capture variations in texture, lighting, and viewpoint. The trained model was deployed on an NVIDIA Jetson Nano, requiring additional optimization to meet real-time inference constraints on embedded hardware.
To enable spatial reasoning beyond 2D detection, I integrated an Intel RealSense RGB-D camera and leveraged depth measurements to support 3D surface analysis. The final system achieved 91.9% detection precision and demonstrated stable real-time performance, making it suitable for on-device inspection tasks in constrained environments.