AAI_2025_Capstone_Chronicles_Combined

15

Draw, Detect, Navigate ​

​

​

​

​

​

​

​

​

Figure 4

Nano Models accuracy on 1,000 images, 5241 doodles within the images

Figure 5

Small Models accuracy on 1,000 images, 5241 doodles within the images

This best performing model was exported in the Open Neural Network Exchange (ONNX)

file formatted and imported into a Unity 6 project using the AI support package Sentis. Using

C#, live frames from a webcam video feed were converted to tensors, then fed to the model.

Model output was subsequently copied back to the CPU, and parsed to visualize bounding

boxes. The simple webcam approach did not have the Simultaneous Localization and Mapping

(SLAM) or depth data available, so for paper surface alignment, the OpenCV package for ArUco

marker tracker was used. Using this approach, drawings are detected in real time on every

frame, with an average Frames Per Second (FPS) of 28. Upon user button press, the centers of

42