EgoLog: Ego-Centric Fine-Grained Daily Log with Ubiquitous Wearables
Despite advances in human activity recognition (HAR) with different modalities, a precise, robust, and accurate daily log system is not yet available. Current solutions primarily rely on controlled, lab-based data collection, which limits their real-world applicability. The challenges towards a fine-grained daily log are 1) contextual awareness, 2) spatial awareness, and 3) effective fusion of multi-modal sensor data. To solve them, we propose EgoLog, which integrates effective audio-IMU fusion for daily log with ubiquitous wearables. Our approach first fuses audio and IMU data from two perspectives: temporal understanding and spatial understanding. We extract scenario-level features and aggregate them in the time dimension, while using motion compensation to enhance the performance of sound source localization. The knowledge obtained from these steps is then integrated into a multi-modal HAR framework. Here, the scenario provides prior knowledge, and the spatial location helps differentiate the user from the background. Furthermore, we integrate a LLM to enhance scenario recognition through logical reasoning. The knowledge derived from the LLM is subsequently transferred back to the local device to enable efficient, on-device inference. Evaluated on both public and self-collected dataset, EgoLog achieves effective multimodal fusion for both activity and scenraio recognition, outperforms the baseline by 12% and 15%, respectively.