Interactive Multi-Touch Gesture Recognition using Circular Measurements
Ruben Balcazar, Francisco R. Ortega, Katherine Tarre, Armando Barreto, Mark Weiss, and Naphtali D. Rishe
CircGR is a multi-touch non-symbolic gesture recognition algorithm, which utilizes circular statistic measures to implement linearithmic (O(n log n)) template-based matching. CircGR provides a solution to gesture designers, which allows for building complex multi-touch gestures with high- confidence accuracy. We demonstrated the algorithm and described a user study with 60 subjects and over 12,000 gestures collected for an original gesture set of 36. The accuracy is over 99% with the Matthews correlation coefficient of 0.95. In addition, early gesture detection was successful in CircGR as well.Contributions
CircGR is a novel way to recognize non-symbolic gestures for interactive (i.e., real-time: the output of the overall sequence of processes appears to be generated without any appreciable delay with respect to the overall input) systems using circular measures as a way to represent gesture templates. The contributions detailed in this paper include: (1) demonstration of a linearithmic algorithm for multi-touch detection with a high recognition accuracy (ACC = 99% and MCC = 0.95); (2) demonstration of early-detection of gestures (with- out the need for gesture completion) with ACC=99% and MCC = 86% for the first window (64 points); (3) demonstration of the need for only one user or developer-provided training sample per template; and (4) a 60-subject experiment with over 12,000 gestures collected for a gesture set of 36. The novelty of CircGR is the simplicity seen in previous Dollar ($) family algorithms, coupled with an improved running time and the ability to recognize a wide variety of non-symbolic gestures for multi-touch displays. As with the $ Family, CircGR provides the ability of gesture designers to create gestures by example and quickly prototype gesture sets. Note that CircGR can run in linear time if sorting is not used during the classification, with minimal decrease in accuracy.Paper
C++ API version coming 2018!Contact
Support provided by the National Science Foundation: I/UCRC IIP-1338922, III-Large IIS-1213026, MRI CNS- 1429345, MRI CNS-1532061, MRI CNS-1532061, MRI CNS-1429345, RAPID CNS-1507611, DUE-1643965. U.S. DOT Grant ARI73. We acknowledge Lukas Borges, Vesna Babarogic, Alain Galvan, and Jonathan Bernal. Finally, Jake Wobbrock, Lisa Anthony, and Radu-Daniel Vatavu for the many email discussions about multi-touch gesture recognition.