Motif Discovery and Feature Selection for CRF-based Activity Recognition

Abstract

Abstract-Due to their ability to model sequential data without making unnecessary independence assumptions, conditional random fields (CRFs) have become an increasingly popular discriminative model for human activity recognition. However, how to represent signal sensor data to achieve the best classification performance within a CRF model is not obvious. This paper presents a framework for extracting motif features for CRF-based classification of IMU (inertial measurement unit) data. To do this, we convert the signal data into a set of motifs, approximately repeated symbolic subsequences, for each dimension of IMU data. These motifs leverage structure in the data and serve as the basis to generate a large candidate set of features from the multi-dimensional raw data. By measuring reductions in the conditional log-likelihood error of the training samples, we can select features and train a CRF classifier to recognize human activities. An evaluation of our classifier on the CMU Multi-Modal Activity Database reveals that it outperforms the CRF-classifier trained on the raw features as well as other standard classifiers used in prior work.

Human Activity Recognition in AAL Environments Using Random Projections

“…Random projections have been used in the HAR domain for data dimensionality reduction in activity recognition from noisy videos [69], feature compression for head pose estimation [70], and feature selection for activity motif discovery [71]. The advantages of random projections are the simplicity of their implementation and their scalability, robustness to noise, and low computational complexity: constructing the random matrix R and projecting the d × N data matrix into k dimensions are of order O ( dkN ).…”

Com

“…The supervised activity recognition methods range from simple methods, such as naive Bayes [Brdiczka et al 2005] based on sensor events independence assumption, to more recent and sophisticated methods, such as conditional random fields Zhao et al 2010], which model the sensor events as probabilistic sequences. Other notable supervised methods include decision trees [Maurer et al 2006], neural networks [Mozer et al 1998], Markov models [Liao et al 2005], and dynamic Bayes networks [Inomata et al 2009].…”

“…Recently, a few unsupervised methods have been proposed to tackle the data annotation problem, such as the frequent sensor mining method [Gu et al 2009], simultaneous frequent-periodic pattern mining method [Rashidi and Cook 2009], episode discovery ], activity modeling based on low-dimensional Eigenspaces [Schiele 2006], multidimensional motif discovery [Vahdatpour et al 2009;Zhao et al 2010], mixed discriminative and generative methods [Huynh and Schiele 2006], probabilistic models [Barger et al 2005;Dimitrov et al 2010], and retrieving activities' definitions using Web mining Palmes et al 2010]. Though these methods address the data annotation problem, they consider a simplified version of the problem by ignoring the real-world nature of data, such as its sequential form, possible disruptions (e.g., a phone call in the middle of meal preparation), or variation of the same pattern.…”

“…The activity recognition system recognizes resident's activities from low-level sensor data. A number of supervised activity recognition methods have been proposed in the literature [Mozer et al 1998;Brdiczka et al 2005;Philipose et al 2004;Zhao et al 2010;Maurer et al 2006;Liao et al 2005;Inomata et al 2009]. Supervised methods have the drawback that they require labeled data for training the recognition algorithm.…”

An activity of daily living primitive–based recognition framework for smart homes with discrete sensor data

“…11,12 There are a lot of researchers who study deeply on proposing more powerful classifiers to improve the accuracy of activity recognition. These classifiers mostly derive from several mainstream ones, to name a few, native Bayesian (NB), 13 hidden Markov model (HMM), 14 support vector machine (SVM), 15 and conditional random fields (CRF), 16 and they characterize in different aspects, for example, NB has a remarkable advantage of recognition speed due to its simple calculation principle, while CRF has better contextual power than others. As classifier researchers known, the Weka is a collection of machine learning algorithms for data mining tasks, 17 which is possible to be applied to train activity models and recognize activities.…”