A study on the impact of layout change to knowledge distilled indoor positioning systems

Convolutional neural networks (CNN)-based indoor positioning systems (IPS) have gained significant attention over the past decade due to their ability to provide precise localization accuracy. However, the use of CNNs in these systems comes with a higher computational cost. To tackle this issue, recent studies have introduced knowledge distilled positioning schemes to mitigate the computational burden. Despite the clear possibility of performance degradation due to signal fluctuations, there remains a lack of investigation into the performance of knowledge distilled and CNN based indoor positioning schemes in dynamic indoor environment. To fill this research gap, this paper investigates the practicality of implementing knowledge distilled-based indoor positioning schemes in real-world by analyzing the impact of indoor layout change on these schemes. Results demonstrate that in the case of layout change, the knowledge distilled-based indoor positioning schemes without teaching assistant can still achieve good performance, with an improvement of 11.56% in average positioning error compared to simple CNN model, while taking only 49.05% of the complex CNN model’s execution time. However, the knowledge distilled-based indoor positioning scheme with teaching assistant fails under the same condition as the inclusion of teacher assistant leads to increased error in modeling the received signal strengths (RSS) and locations relationship.