A hierarchical representation for human activity recognition with noisy labels

Abstract

Abstract-Human activity recognition is an essential task for robots to effectively and efficiently interact with the end users. Many machine learning approaches for activity recognition systems have been proposed recently. Most of these methods are built upon a strong assumption that the labels in the training data are noise-free, which is often not realistic. In this paper, we incorporate the uncertainty of labels into a maxmargin learning algorithm, and the algorithm allows the labels to deviate over iterations in order to find a better solution. This is incorporated with a hierarchical approach where we jointly estimate activities at two different levels of granularity. The model is tested on two datasets, i.e., the CAD-120 dataset and the Accompany dataset, and the proposed model shows outperforming results over the state-of-the-art methods.

Nonparametric Bayesian models for unsupervised activity recognition and tracking

“…Sometimes annotators make mistakes which add noise into the labels. Treating these noisy labels as the ground truth is typically harmful for most learning methods [1]. Steps are sometimes taken to alleviate this labelling bias, e.g.…”

“…[18] suggested a method that models each label as a multinomial distribution rather than deterministic. In [1], the authors treat all of the labels as noisy data, and add minor probability mass to incorrect labels enabling the model to converge to a better representation of the actions. Hence, with this in mind we use the labels with caution, in inferring statistical properties of the observations.…”

“…Being able to determine the mode of locomotion of a human user or the system's place in its operating environment, is necessary for effective use of robotic systems for everyday usage; from aiding elderly users with their daily activities of living to providing accurate assistance to the user in an industrial and/or home setting. Although many previous works have focused on this topic, recognising complex activities endures as a challenging task which is still an open problem [1]. We contribute a fully unsupervised approach to the problem of locomotion and activity recognition, through the use of Bayesian nonparametric state-space models (SSM) that use probabilistic programming (PPS) general-purpose inference [2].…”