Naturalistic driving studies rely on human data reductionists to manually review and annotate driving behaviors. This work is time-consuming, and algorithms that could scan and categorize video data could make the data reduction process faster and more efficient. This report describes research to develop pose estimation methods that can be applied to drivers in naturalistic settings. Three methods were explored: (1) a depth-sensor-based pose estimation; (2) a deformable parts-based model; and (3) a tiny-image-based driver activity classifier. The tiny-image-based approach was chosen as the final solution and tested using the VTTIMLP01 dataset, a collection of about 80,000 images from 25 participants in naturalistic driving and simulated naturalistic driving conditions. The model was applied to approximately 50,000 images from the dataset covering seven activity classes: Eating/Drinking, Talking, Visor, Center Stack, Texting, One Hand on the Wheel, and Both Hands on the Wheel. The model, without any aspect ratio changes to the input image, was able to predict the activity classes with an overall 70% accuracy. To obtain better accuracies for individual activity classes, a separate model was built for each class, which resulted in a model with an overall accuracy of 74%. The Texting class had the poorest class accuracy (56%) due to the foreshortening effect on the limbs in the given camera angle. The One Hand on the Wheel class had the best class accuracy (96%).