This study introduces a novel approach to leverage low-power, low-resolution infrared sensors for detailed people tracking in manufacturing settings. We curated a dataset including a diverse range of interactions labeled for multiple-person localization and social distance violation tasks. Our methodology uses a combination of convolutional and recurrent neural networks to interpret spatiotemporal data. We demonstrate the capability of the novel image segmentation approach for human localization where we achieve 97.5 percent image-level accuracy. Also, we highlight the importance of interpolation and convolutional kernel selection for social distance tasks where we achieve 91 percent macro-averaged accuracy in 4 class scenarios.