AIO: Enhancing Generalization of Inertial Odometry to Unseen Scenarios with Active Learning

Inertial odometry (IO) emerges as a promising direction for location estimation lately, mainly due to the proliferation of inertial measurement units in mobile devices. Combined with the deep neural networks, IO learns the non-linear mapping between inertial readings and trajectory. Despite the high accuracy, IO usually requires extensive fine-grained datasets for model training, which is usually prohibitively-costly to collect, leading to compromised general-ity to new pedestrians, holding positions and walk patterns. To address this, we propose AIO, an active learning-based inertial odometry approach. Using a few selected inertial sequences, we train the model in a incremental manner and instruct the model to learn and generalize to unseen inertial sequences. Instead of tedious labeling of all inertial sequences, we propose a feedback-based selection paradigm, where we leverage both geometrical and temporal metrics to select the most distinguishing sequence to increase the information of the model. Finally, we propose a transformer-based inertial encoder, which learns more accurate embeddings of inertial sequences. We have implemented the proposed approach and conducted extensive evaluations with major datasets and comparison schemes. Experimental results demonstrate that the proposed approach outperforms competing approaches in unseen scenarios with only one labeled inertial sequence.