Perception & User Data Sets-Quality of Life Technology Center - Carnegie Mellon University

Perception & User Data Sets

The project addresses basic research in areas such as sensing, logging and analyzing contextual data, modeling ranging from statistical to inferential and systematic evaluation of the quality and relevance of various sensor modalities. The resulting models are used for generating prediction from sensor data and identifying intent for virtual coaches, therefore contributing to increasing a person’s independence.

We have shown that low fidelity sensors (such as accelerometers) and our machine learning algorithms can be used to recognize propulsion patterns for manual wheelchair users. This provides insight into how the users should perform these actions, as well as guidance to perform them correctly, namely, it can aid in the training of manual wheelchair users to avoid the more strenuous propulsion techniques, which can lead to repetitive use injuries.

We conducted the first study of classifying wheelchair propulsion patterns using low-fidelity, body-worn sensors. We developed a manual wheelchair propulsion classification system which recognizes different patterns using our wrist mounted eWatch’s accelerometer. Four classic propulsion patterns have been identified by a limited user study, which are semicircular, single loop over, double loop over and arcing. Of these, the recommended propulsion pattern is semi-circular, because the strokes have lower cadence and higher stroke angle Data was collected using all four propulsion patterns on a variety of surface types. The results of two common machine learning algorithms, k-Nearest Neighbor (kNN) and Support Vector Machines (SVM) with a Radial Basis Function (RBF) kernel were compared. Accuracies of over 90% were achievable even with a simple classifier such as k-Nearest Neighbor (kNN). Hence, it is demonstrated that an accurate propulsion pattern classifier can be created using only a single wrist mounted accelerometer. It was also noted that the higher the resistance of the surface traversed, the higher were achievable accuracies.

The eWatch was used to collect propulsion data from a subject over several surfaces. In the light that this data will be useful for other experiments, we have documented the format of this data as part of the Grand Challenge Data Gathering.

Project Team

  • Dan Siewiorek, Lead
  • Byron Boots
  • Anind Dey
  • Brian French
  • Jennifer Mankoff
  • Asim Smailagic
  • Divya Tyamagundlu
  • Brian Ziebart 

 Real-time self-reporting of daily stress using eWatch

eWatch figure