Fine-Grained Activity Recognition Based on Features of Action Subsegments and Incremental Broad Learning

EasyChair Preprint 7218

Abstract

Human activity recognition using MEMS on mobile devices has become one of the most compelling solutions owing to the miniaturization of sen-sors. A crucial challenge is to recognize precisely activities when they are changing. Sliding window is a type of common methods. However, the in-terference of historical data in the sliding window is harmful to insight into changing of actions or uncommon behaviors. This paper proposes a fine-grained activity recognition method and designs a corresponding system farer. It employs features of action subsegments and incremental broad learning to precisely distinguish the alterations of actions and abnor-mal movements. Firstly, farer achieves the accurate segmentation of activi-ties as data preprocessing. A neighborhood extreme value method (NEV) is adopted to avoid the intervention of peaks and valleys of data. Secondly, the current action is partitioned to fine-grained subsegments to elaborately abstract subtle features. We propose a feature extraction technique based on adjacent difference (FETAD), and furthermore reduce its resulting dimen-sion through the complete two-dimensional principal component analysis (C2DPCA). Finally, broad learning theory is employed to construct the ac-tivity recognition model, especially incremental learning for unusual be-haviors. Extensive experiments demonstrate that farer could accurately recognize activities when they abruptly change, and its performance is con-siderable stability. Meanwhile, it can quickly establish a valid incremental model that only needs a short sampling time for special activities. The over-all accuracy of farer is 97.91% with 90.14% for changed activities, which is far superior to the current mainstream methods.

Keyphrases: Action Subsegments, Broad Learning, fine-grained recognition, incremental model

@booklet{EasyChair:7218,
  author    = {Shi Chen and Sheng Wu and Licai Zhu and Hao Yang},
  title     = {Fine-Grained Activity Recognition Based on Features of Action Subsegments and Incremental Broad Learning},
  howpublished = {EasyChair Preprint 7218},
  year      = {EasyChair, 2021}}