Ensemble learning for human activity recognition

Abstract

This paper describes a activity recognition method for Sussex-Huawei Locomotion (SHL) Challenge 2020 by team TDU-BSA. The use of ensemble learning, which combines the outputs of multiple classifiers to produce a single estimation result, improved the accuracy of activity recognition. The ensemble model consists of CNN models and a gradient-boosting model. The objective of SHL Challenge 2020 is that the users of SHL test-set are two different from SHL training-set, and the phone location of SHL test-set is not known to the SHL's participants. Therefore, estimating phone location and the user improved accuracy. SHL test-set's phone location was estimated to be Hips. The user can be estimated from SHL validation-set. The ensemble model was made with all SHL training-set (Only Hips) and 70% of SHL validation-set (Only Hips). In the submission phase, the best F-measure obtained for last 30% SHL validation-set was 84.8%.