Device Agnostic 3D Gesture Recognition using Hidden Markov Models

Anthony Whitehead,

Kaitlyn Fox,

School of InformationTechnology

Carleton University

1125 Colonel By Drive,

Ottawa, ontario, Canada

Comments:

Manoj’s Blog

Franck’s Blog

Summary:

Introduction:

This paper seeks to identify the necessary elements to successfully use Hidden Markov Models for 3D gesture recognition regardless of sensor devices being used. A variety of input sensors in place of a single one results in increasing the alphabet of the model. However, the study has shown that an alphabet larger than 27 becomes computationally too expensive to allow real time interactivity.

Gestures in Training:

The training set was generated from several users wearing an accelerometer on their wrists. They all performed seven gestures.

The Number of HMM States:

A balance between false negatives and false positive needs to be obtained for the model to be viable. 27 states yielded the best performance. The smaller number of the states had a benefit of increased computational performance.

Number of Samples in a Training Set:

250 samples are sufficient as a training set for a system with 27 states.

Culling Training Data:

During training there were some inconsistencies caused by individual interaction with the hardware. To address this the longest and shortest data elements were discarded. A 1.5 standard deviation rule was used to decide which sequences to discard.

Left vs. Right Hand Training:

Gestures performed by the left and right hands were compared. Gestures performed by the left hand were not recognized, this was attributed to a possible difference in the angle of tilt of the accelerometer. There was an almost a 50% decline in performance from the right to the left hand. Interestingly this was not the case with unidirectional gestures.

Results and Conclusion:

Overall:

91.6% correct recognition for gestures in the training set.

86.4% correct recognition for gestures not in the training set.

Discussion:

The simplicity and adaptability of a wide selection of input sensors was particularly appealing. The comparison between right and left handedness is interesting. There is no mention of the difference in dexterity differences demonstrated by most subjects. A more representative evaluation may have been to test the performance of right handed trained models on left hand gestures performed by left handed subjects.