Enhancing Locomotion Mode Recognition Accuracy: A Deep Learning Model for Noisy Label Handling

Abstract

The availability of various sensors in smartphones has made it easier to collect human locomotion data. This data is useful for recognizing different locomotion modes namely walking, cycling and traveling. The correct identification of locomotion modes helps in transportation planning, traffic analysis and travel cost estimation. In most cases, these labels are assigned through crowdsourcing or web-based queries. This paper presents a deep learning-based approach for locomotion mode recognition in the presence of noisy labels. The proposed method builds an ensemble model consisting of three sub-models. The different models were created for processing training data with varying degrees of noise contaminations. A conventional deep learning system obtains spatial-temporal information from basic sensor inputs. The model maintains reliable performance within conditions of minimal label noise. A noise adaptive model contains a dynamic loss function which adjusts its operation according to noise intensity levels. The model shows success in dealing with intense noise levels. The proposed approach showed its effectiveness through experiments conducted on both newly obtained datasets and pre-existing available datasets. This paper resolves the problems with noisy labels which occur in locomotion mode recognition tasks. The ensemble model achieves recognition success through its integration of conventional, adaptive and corrective techniques.