Fine-Tuning AlexNet for Bed Occupancy Detection in Low-Resolution Thermal Sensor Images

Low resolution thermal sensing technology is particularly well suited to activity monitoring in the bedroom environment due its ability to operate irrespective of lighting conditions and its privacy conserving nature. This paper investigates the application of transfer learning with AlexNet to classify bed occupancy in temperature image data. Problem specific tailor-made CNNs with 3 or 4 layers are generally developed and trained from scratch to classify low resolution thermal sensor image data. To date, no research has evaluated the use of a pre-trained or fine-tuned CNN on low resolution thermal sensor image data. Transfer learning is particularly useful for specialized tasks, such as detecting bed occupancy within thermal sensor images, as large training datasets are not readily available. In this paper, 3 different fine-tuning configurations of the AlexNet architecture are evaluated. The networks are trained on a balanced two-class dataset of over 90,000 images and tested using the leave one subject out validation method. In this study, the best performing network had 3 learnable layers and achieved an accuracy of 0.973 on greyscale images with a temperature resolution of 220 x 220.