[Paper Review] Recognizing Activities and Spatial Context Using Wearable Sensors

Amarnag Subramanya, Alvin Raj+2 more|arXiv (Cornell University)|Jun 27, 2012

Context-Aware Activity Recognition SystemsComputer Science17 references78 citations

TL;DR

This paper proposes a dynamic graphical model that jointly infers human activities and spatial context using asynchronous wearable sensor and GPS data. By fusing multi-modal sensor inputs through particle filtering and pruned exact inference, the model achieves higher accuracy than using GPS or sensors alone, while minimizing hardware load for improved comfort and battery life.

ABSTRACT

We introduce a new dynamic model with the capability of recognizing both activities that an individual is performing as well as where that ndividual is located. Our model is novel in that it utilizes a dynamic graphical model to jointly estimate both activity and spatial context over time based on the simultaneous use of asynchronous observations consisting of GPS measurements, and measurements from a small mountable sensor board. Joint inference is quite desirable as it has the ability to improve accuracy of the model. A key goal, however, in designing our overall system is to be able to perform accurate inference decisions while minimizing the amount of hardware an individual must wear. This minimization leads to greater comfort and flexibility, decreased power requirements and therefore increased battery life, and reduced cost. We show results indicating that our joint measurement model outperforms measurements from either the sensor board or GPS alone, using two types of probabilistic inference procedures, namely particle filtering and pruned exact inference.

Motivation & Objective

• To develop a system that simultaneously recognizes human activities and spatial context using wearable sensors.
• To reduce hardware burden by minimizing the number of sensors worn, enhancing user comfort and battery efficiency.
• To improve recognition accuracy by jointly modeling activity and spatial context using asynchronous sensor and GPS data.
• To evaluate the performance of the joint model against standalone GPS or sensor-based approaches.
• To demonstrate the feasibility of real-time inference using efficient probabilistic techniques like particle filtering and pruned exact inference.

Proposed method

• The system employs a dynamic graphical model to model the joint probability of activities and spatial context over time.
• It fuses asynchronous observations from a wearable sensor board (accelerometer, gyroscope) and GPS measurements.
• Particle filtering is used for approximate inference to handle the continuous state space and non-linear dynamics.
• Pruned exact inference is applied to improve computational efficiency and accuracy in certain configurations.
• The model uses conditional probability distributions to relate sensor inputs to activity and location states.
• A hidden Markov model framework is extended to include spatial context as a latent variable, enabling joint estimation.

Experimental results

Research questions

• RQ1 Can joint inference of activity and spatial context improve recognition accuracy compared to using sensors or GPS alone?
• RQ2 How does the fusion of asynchronous sensor and GPS data affect the performance of activity and location recognition?
• RQ3 To what extent can hardware load be reduced while maintaining high recognition accuracy?
• RQ4 How do particle filtering and pruned exact inference compare in terms of accuracy and efficiency for this joint inference task?
• RQ5 Can a dynamic graphical model effectively model the temporal dependencies between activities and spatial contexts?

Key findings

• The joint measurement model significantly outperforms models using only GPS or only sensor data in activity and spatial context recognition.
• Particle filtering and pruned exact inference both enable accurate inference, with pruned exact inference showing higher accuracy in controlled evaluations.
• The system achieves improved recognition accuracy while minimizing the number of sensors worn, enhancing user comfort and battery life.
• The fusion of asynchronous sensor and GPS data leads to more robust and consistent state estimation over time.
• The dynamic graphical model effectively captures temporal dependencies between activities and spatial contexts, improving overall inference quality.
• Empirical results from the UAI 2006 conference demonstrate the effectiveness of the joint inference approach in real-world settings.