Toward Designing an Attentive Deep Trajectory Predictor Based on Bluetooth Low Energy Signal

In this study, a novel attentive deep trajectory predictor is proposed for personal key (PK) localization problem in a Bluetooth low energy (BLE) network. This model has a unique sparseness design enlightened by the physical nature of the PK localization problem. Moreover, a set of geometrically inspired embedding losses are proposed to enhance model's generalization ability on different BLE anchor layout. Finally, the trained model with tiny footprint is deployed in a low-end vehicle processor. Intensive tests and carefully designed ablation studies are conducted to prove the robustness and effectiveness of the model.