Data augmentation for Long Short-Term Memory Neural Networks in trajectory prediction of Indoor Person Localization from infrared sensors =

There are many applications for indoor personnel tracking and localization using infrared thermal sensors, such as health monitoring and indoor security surveillance. This method is the best suitable for tagless localization of human bodies, adapting to different users and different scenarios. Since Infrared sensors are more affordable than other sensors used to detect and track people and their activities, infrared sensors are currently widely used in the Internet of Things. This thesis evaluates the Long Short-Term Memory neural networks for predicting data from a low-resolution 16-pixel thermopile sensor data for indoor localization and tracking, improving robustness and reliability by adding unrelated noise to the sensor data. This noise addition is a type of data augmentation that results in the production of more data. Currently, the Long Short-Term Memory neural networks has been proposed for their excellent performance in a variety of tasks such as speech recognition and machine translation. By training neural network models on significantly more data, data augmentation improves the ability of the models to generalize what they have learned to the new data and the unseen data. In this thesis, the model results obtained with augmented data are compared to the model baseline initially generated without augmentation techniques. Data augmentation noise is then added in varying amounts compared to the infrared sensor signal variance between human presence and absence. We have four independent datasets, one of which was used mainly for training and partly for testing, and the other three only for testing. The average mean square error across all sets is used to calculate the generalization quality, while the learning curves are used to determine the learning quality. The noise amplitude is varied, trying to find the optimal improvement over the baseline for generalization and learning quality. The inference performance and learning quality change differently for each data set with the noise amplitude such as one set shows a noticeable improvement while the others do not. To avoid false conclusions, the regions of interest with lower mean square error values would be examined in detail. If the overall performance of the model improves for some noise amplitudes, the refined ranges would be examined to see if the model has better generalization for all sets. To test the generalization ability of the model's inference. Initially, the data augmented by solely adding white noise which is frequently obtained from various sources, followed by augmented the data with only brown noise which is characteristic of low-frequency variations. The amplitude of each noise comprised 0.01% to 163.84% of the signal variance between person presence and person absence. The performance of the model is found to be improved best in the range [11.6%; 36.1%] of the white noise range and [0.1%; 0.59%] of the brown noise amplitude. Then, when two types noises are augmented together by combination, the best improvement can be found when the white noise amplitude is 17% and the brown noise amplitude is 0.32%.