The sensorimotor cortex plays a fundamental role for animals and humans to navigate and act in complex environments. In conventional experimental approaches animals perform highly trained tasks in a trial-repetitive manner with their movements limited to ensure reproducibility of the same or similar trial conditions. These conditions limit the study of ecologically relevant foraging behavior, particularly those involving full-body movements. However, recording from freely moving animals poses significant challenges for computational analysis due to the uniqueness of each trial and the complex neuronal signals from full-body motion. In this work, we take a step towards addressing these challenges by analyzing data from freely moving macaques in an ecologically relevant environment. Specifically, we analyze the spatial frame of reference in which neurons in the primary motor cortex (M1) and dorsal premotor cortex (PMd) encode information. To this end, we train a deep learning model that predicts neural activity as a function of full-body motion tracked via keypoints from videos. We subsequently remove spatial information from the keypoints to quantify the effect on the model's predictive performance. Specifically, we train models on three different representations of the keypoints: 1) an allocentric coordinate system, 2) a centered coordinate system that preserves global direction but removes information about global location, and 3) an egocentric coordinate system that removes global information about both location and direction. We expect to see a drop in the model's performance if we remove important information encoded by the electrode channel. However, we do not find a statistically significant drop in any of the brain areas, suggesting that neurons in M1 and PMd encode information in a predominantly egocentric way. Overall, our work provides a novel deep learning method to studying neuronal encoding in freely-behaving animals, opening new possibilities for understanding the sensorimotor cortex.