An Unsupervised Approach for Automatic Activity Recognition based on Hidden Markov Model Regression (1312.6965v1)

Abstract: Using supervised machine learning approaches to recognize human activities from on-body wearable accelerometers generally requires a large amount of labelled data. When ground truth information is not available, too expensive, time consuming or difficult to collect, one has to rely on unsupervised approaches. This paper presents a new unsupervised approach for human activity recognition from raw acceleration data measured using inertial wearable sensors. The proposed method is based upon joint segmentation of multidimensional time series using a Hidden Markov Model (HMM) in a multiple regression context. The model is learned in an unsupervised framework using the Expectation-Maximization (EM) algorithm where no activity labels are needed. The proposed method takes into account the sequential appearance of the data. It is therefore adapted for the temporal acceleration data to accurately detect the activities. It allows both segmentation and classification of the human activities. Experimental results are provided to demonstrate the efficiency of the proposed approach with respect to standard supervised and unsupervised classification approaches

An Examination of Unsupervised Human Activity Recognition via Hidden Markov Model Regression

The paper presented by D. Trabelsi et al. introduces an innovative unsupervised methodology for human activity recognition leveraging acceleration data from wearable sensors. The core approach utilizes the Hidden Markov Model Regression (HMMR) framework in a multivariate setting to facilitate the detection of activities without necessitating labeled datasets.

Overview of Methodology

The challenge addressed by the research is the automatic recognition of human activities using data captured by accelerometers embedded in wearable devices. Traditional supervised models demand extensive labeled data, which is often expensive and impractical to gather in realistic settings. In contrast, the proposed method develops an unsupervised framework that circumvents this requirement by applying HMM within a regression context.

The approach formulates the activity recognition task as a segmentation problem of multidimensional time series data. Specifically, the activity sequences are modeled using a piecewise representation where a Hidden Markov Chain (HMC) guides transitions between different activity regimes. Each regime is described by a polynomial model learned from the data using the Expectation-Maximization (EM) algorithm, which processes the temporal structure inherent in human activity data efficiently.

Data and Experimental Framework

Data was collected through an experimental setup involving six subjects who performed a range of everyday activities. Three tri-axial MTx inertial sensors were employed—located on the chest, right thigh, and left ankle. This configuration ensured comprehensive coverage for discerning upper and lower-body activities. The defined activities included a mix of static and dynamic actions such as walking, sitting, lying, and various transitions between these states.

Experimental Results and Evaluation

The results from applying the MHMMR approach were robust, with an impressive classification accuracy of 91.4%, a promising achievement in an unsupervised setting. The segmentation method effectively delineated between different activities, even though it encountered some challenges in classifying transitions between activities—a common difficulty due to their brief durations. When compared to baseline models, such as $k$-Means clustering and Gaussian Mixture Models (GMM), the MHMMR significantly outperformed in terms of correct classification rate, suggesting its superiority for handling sequential and temporal data.

Additionally, a comparative analysis with supervised models revealed that the proposed approach competes well, though techniques like $k$-NN delivered marginally better performance due to their supervised nature. However, these supervised techniques necessitate comprehensive labeled datasets, which pose a limitation the unsupervised MHMMR approach overcomes.

Implications and Future Directions

This paper contributes significantly to unsupervised learning methods in the field of human activity recognition, providing a method that does not over-rely on labeled data. Future work could explore the integration of Bayesian models to adaptively handle model complexity and further delve into non-parametric approaches that inherently adjust to various activities without predefined structures.

In terms of application, this method holds potential in health-monitoring systems for elderly care or rehabilitation contexts, where continuous, label-free monitoring could offer significant benefits. Furthermore, expanding this model to incorporate data fusion techniques with other sensor types could offer richer profiles for accurate activity recognitions and reduced computational burden.

Overall, the paper proposes a viable alternative for contexts where supervised learning is infeasible, setting a foundation for more dynamic and adaptable activity recognition systems in real-world applications.