Research On Continuous Letter Trajectroy Hand Gesture Recognition And Its Application

Continuous trajectory gesture recognition plays an important role in Human Computer Interaction(HCI),the natural and precise continuous trajectory recognition approach can improve the efficiency of HCI,which is widly used in medical and robotic field.Traditional continuous trajectory have some drawbacks such as unnatural trajecotry segmentation,poor robustness for different complex trajectory.The main contributions of this paper are as follow:A model based continuous trajectory segmentation approach is proposed to make segment process natural.The main ideal of which is segmenting the trajectory by detecting the writing speed down area in the end of trajectory.We first detect the modellike trajectory using DTW,then,detecting the writing speed down area from time series at equal intervals to filter the noise caused by writing.The proposed approach can segment continuous trajectory naturally and accurately,which is more efficient than traditional segmentation approach.In order to improve the robustness for different complex trajectory,a structured dynamic time warping approach is proposed,which uses both orientation feature and position feature to make up the shortage of each other.To making position feature robust for scale changing,PCA approach is used to normalize it in principal direction.Besides,we use structure information and weight strategy to improve traditional DTW,compared with it,our approach have better recognition performance and robustness.Multi-templates are used to reflect one class in our approach to over come different writing styles.We propose a templates selecting approach based on K-means method and a parameters optimizing approach based on GA to find representative templates.Moreover,we propose a single fingertip tracking approach to reduce the noise caused by interaction in multi-degree of freedom.The recognizing results of different people are close,which proves our approach is robust to different writing styles.The proposed continuos trajectory recognition approach based on structure dynamic time warping is mainly evaluated on continuous letter trajectory database and the F score of it is 86.6%,which is better than traditional dynamic time warping.Moreover,the proposed approach is evaluated on continuous digital database too,and the recognition rate of it is 95.8%.Experiment results show that the proposed approach can accurately and precisely segment trajectory on real-time and recognize trajectory with different complexity.Furthermore,we apply the proposed approach in medical field and point out its good application prospect.