Investigation of Heterogeneity Sources for Occupational Task Recognition via Transfer Learning

Abstract

Human activity recognition has been extensively used for the classification of occupational tasks. Existing activity recognition approaches perform well when training and testing data follow an identical distribution. However, in the real world, this condition may be violated due to existing heterogeneities among training and testing data, which results in degradation of classification performance. This study aims to investigate the impact of four heterogeneity sources, cross-sensor, cross-subject, joint cross-sensor and cross-subject, and cross-scenario heterogeneities, on classification performance. To that end, two experiments called separate task scenario and mixed task scenario were conducted to simulate tasks of electrical line workers under various heterogeneity sources. Furthermore, a support vector machine classifier equipped with domain adaptation was used to classify the tasks and benchmarked against a standard support vector machine baseline. Our results demonstrated that the support vector machine equipped with domain adaptation outperformed the baseline for cross-sensor, joint cross-subject and cross-sensor, and cross-subject cases, while the performance of support vector machine equipped with domain adaptation was not better than that of the baseline for cross-scenario case. Therefore, it is of great importance to investigate the impact of heterogeneity sources on classification performance and if needed, leverage domain adaptation methods to improve the performance.

Comparing Cross-Subject Performance on Human Activities Recognition Using Learning Models

“…With the growth of deep learning research, recent HAR studies focus on the improvement of recognition accuracy using complex deep architectures or transfer learning [53]- [56] rather than traditional solutions [4], [5], [14]. However, some studies [64], [65] discovered the phenomenon that the traditional solutions outperform deep methods under the same metric on HAR, and the reason remained unclear.…”

“…(2) Unsupervised domain adaptation aligns the feature distributions between source and target domains by means of distance minimization [53]- [56] or generative adversarial networks (GAN) [57]- [60]. Hosseini et al [53] designed a BLSTM to extract representative features and minimize confusion between source and target domains through maximum mean discrepancy loss.…”