Activity Recognizer

Type: Diploma thesis
Student: Thomas Brügger

This diploma thesis treats gesture and activity recognition based on a commercial inertial sensor system Xsens (http://www.xsens.com). The thesis proposes a framework of modular elements of a gesture recognition process. This framework starts with acquiring the data, which includes a definition of a test gesture set and also a definition of a standard data acquisition procedure.

The next element of the framework is the segmentation of the continuous orientation data stream produced by the sensors. The first step of the segmentation is to derive the angle signals in order to obtain the corresponding angle velocity signals. Then the segmentation procedure detects the zero crossings in the velocity signals marks them as starting points of segments and saves them in a segment starting points vector.

Once the signal has been segmented, a set of features (such as number of segments or the length of all segments of a single gesture) is evaluated on a sequence of training gestures.

The core of the gesture recognition framework is a new algorithm which uses a simple and fast minimum distance classifying scheme. The algorith starts a search process from each segment starting point to a adjustable segment depth. If the sum of euclidian distances to the classes of target gestures at a segment end point goes below a class specific threshold value, the corresponding start and end segments are saved in a gesture candidates list. This candidate list can also be used as input for a more sophisticated classifying process.

The last element in the thesis' framework is a method for statistical analysis which is based on comparing the results of the search algorithm with the labelling data that is recorded in parallel with the gesture sequences.