Analysis and Selection of Classifiers for Gesture Characteristics Recognition Based on MGRS Framework
Justin van Heek, Gideon Woo, Jack Park, Herbert H. Tsang · 2019
Human gesture is a complex motion that can be encoded with a multitude of information. Gesture recognition is the process of decoding and classifying this information. Currently, there are various computational intelligence techniques that have been developed to recognize gestures. Due to the vast quantity of algorithms available, it is difficult to understand which algorithms perform best for a given gesture recognition problem. Current studies in gesture recognition usually focus on a singular gesture characteristic for recognition leaving a gap in knowledge on the role multiple characteristics play in gesture recognition. In this study, various classification a lgorithms are evaluated by classifying the beat pattern and the articulation of a conducting gesture, two very distinct characteristics of the same gesture. The results of this study show that all of the top performing algorithms for both characteristics arc part of the neural network family, specifically the recurrent neural networks. It was also discovered that some algorithms had significantly different ranges of accuracy depending on the characteristic of the gesture. On the other hand, other algorithms showed less than a 5% difference between the two characteristics. From these results we determine that neural network algorithms should be considered for any gesture recognition problem regardless of its characteristics due to their relatively low variance between characteristics. The existence of potentially significant variance in results between different characteristics for other algorithms reveal that the characteristics of a gesture play an important part in determining the performance of a given algorithm.