Optimized limited memory and warping LCSS for online gesture recognition or overlearning?

In this paper, we present and evaluate a new algorithm for online gesture recognition in noisy streams. This technique relies upon the proposed LM-WLCSS (Limited Memory and Warping LCSS) algorithm that has demonstrated its efficiency on gesture recognition. This new method involves a quantization step (via the KMeans clustering algorithm). This transforms new data to a finite set. In this way, each new sample can be compared to several templates (one per class) and gestures are rejected based on a previously trained rejection threshold. Then, an algorithm, called SearchMax, find a local maximum within a sliding window and output whether or not the gesture has been recognized. In order to resolve conflicts that may occur, another classifier could be completed. As the K-Means clustering algorithm, needs to be initialized with the number of clusters to create, we also introduce a straightforward optimization process. Such an operation also optimizes the window size for the SearchMax algorithm. In order to demonstrate the robustness of our algorithm, an experiment has been performed over two different data sets. However, results on tested data sets are only accurate when training data are used as test data. This may be due to the fact that the method is in an overlearning state.