I led the MobileNet implementation for our four-model comparison study, and prepared the research report, while also supporting teammates in fine-tuning their respective models.

MobileNet was a natural fit for this problem: it’s depthwise separable convolutions keep the parameter count low without a significant accuracy penalty, making it genuinely viable for edge deployment scenarios. I extended the base model by adding two hidden layers with ReLU activation, batch normalization to reduce variance between batches, and dropout to guard against overfitting. I also reworked the training setup to optimize for accuracy rather than the original logarithmic loss, tuned the learning rate to 0.005 for stable convergence, and extended training to 30 epochs with early stopping patience of 20.

A key preprocessing decision was splitting the dataset by driver ID rather than randomly, which prevents data leakage by ensuring the network sees entirely different drivers at test time. I also applied image augmentation (height, width, zoom shifts and 30° rotation) to synthetically expand the training set and improve generalization across lighting conditions and camera angles.