Image Detection Using Nano 33 BLE Arduino Board

Overview

• This project, focuses on object detection, specifically differentiating between stop signs and non-stop signs.
• Utilizing the Arduino Nano 33 BLE and the OV7675 Camera Module, the project aims to train a binary classification model and deploy it onto an embedded board.
• The dataset combines ten self-derived images and a comprehensive online dataset from Kaggle, tailored to the project's requirements.

Model Architecture

• The model architecture was refined through experimentation, with a cap at 300k parameters.
• The final model inputs are grayscale images of size 96x96 pixels.
• Various architectures were tested, including MobileNetV1 and models with different parameters and activation layers, leading to the selection of the most effective model in terms of accuracy and parameter count.

Data and Training

• Custom images include diverse shots of stop signs and similar signs under different conditions to challenge the model's accuracy.
• Data augmentation techniques were applied to increase the diversity of the training data.
• The dataset was segregated into training, testing, and validation sets, with a focus on achieving a balance between stop and non-stop sign images.
• Detailed account of the training process, model choices, and considerations regarding overfitting and dataset composition.

Results

• The model achieved a training accuracy of 87.06%, validation accuracy of 90.32%, and a test accuracy of 87.05%.
• The model demonstrated a low false rejection rate (0.15%) and false positive rate (0.04%).
• Detailed insights into the model's performance, including its ability to correctly identify images and limitations in detection under certain conditions.
• Graphical representations of loss and accuracy over training epochs are included.