3D Imaging for Behavior Classification, 10-R8221
Principal Investigator
Chris Lewis
Inclusive Dates: 04/01/11 – 10/01/12
Background — This research developed an automated behavior recognition capability, which uses a very low-cost, 3D color sensor for observing the motion of people. The system uses a variety of state-of-the-art machine learning techniques to estimate which of the trained behaviors is being performed. Several training tools were also developed which allow the system to be easily customized for a variety of applications. A novel feature derived from raw motion measurements was developed and shown to discriminate well between exercise behaviors. This feature, called a Motron, is constructed from natural cluster centers in data vectors containing position and velocity measurements of the subject. A new clustering algorithm was also developed and shown to be useful for both analysis and for accurately modeling sampled data.
Approach — The techniques were implemented under ROS (Robot Operating System), which is an open architecture, publish-subscribe system that integrates drivers for common sensors and machine learning tools into a convenient development environment. ROS nodes were developed for training classifiers, analyzing clusters in data and estimating behaviors in real time.
Accomplishments — A novel motion descriptor, called a Motron, was developed that is formed from natural clusters in pose measurements. Histograms of observed Motrons over a time window were shown to be both salient and computationally inexpensive features for classifying behaviors in real time. In addition, a novel cluster analysis algorithm was developed which automatically determines both the number of clusters and models for those clusters in arbitrary high-dimensional data.
