Proceedings of the 28th International BCS Human Computer Interaction Conference (HCI 2014) (HCI)
BCS Human Computer Interaction Conference (HCI 2014)
9 - 12 September 2014
This paper systematically explores the capabilities of different forms of Dynamic Time Warping (DTW) algorithms and their parameter configurations in recognising whole-of-body gestures. The standard DTW (SDTW) (Sakoe and Chiba 1978), globally feature weighted DTW (Reyes et al. 2011) and locally feature weighted DTW (Arici et al. 2013) algorithms are particularly considered, while an enhanced version of the globally feature weighted DTW (EDTW) algorithm is presented. A wide range of configurable parameters: distance measures (Euclidean and Mahalanobis), combination of features (Cartesian velocity, angular velocity and acceleration), combinations of skeletal elements, reference signal count and k-nearest neighbour count are tested in order to understand the impact on final recognition accuracies. The study is conducted by collecting gesturing data from10 participants for 9 differentwhole-of-body gesture commands. The results suggest that the proposed enhanced version of the globally feature weighted DTW algorithm performs significantly better than the other DTW algorithms. Given sufficient training data this study suggests that the Mahalanobis distance has the capability to better differentiate certain gestures compared to the Euclidean distance. Out of the features, Cartesian velocity combined with angular velocity provides the highest gesture discriminant capability while acceleration provides the lowest. When highly informative and stable skeletal elements are selected, the overall performance gain obtained by adding extra skeletal data is marginal. Also the recognition accuracies are sensitive to the reference signal count and the KNN percentage. Additionally, the presented results summarise the unique capabilities of certain configurations over others, highlighting the importance of selecting the appropriate DTW algorithm and its configurations to achieve optimal gesture recognition performances.