| Popis: |
Abstract When it comes to the studies concerning human-computer interaction, human posture recognition which was established on the basis of Kinect is widely acknowledged as a vital studying field. However, there exist some drawbacks in respect of the studying methods nowadays, for instance, limitations, insignificantly postures which can be recognized as well as the recognition rate which is relatively low. This study proposed a brand new approach which is hybrid in order to recognize human postures. The approach synthetically used depth data, skeleton data, knowledge of anthropometry, and backpropagation neural network (BPNN). First, the ratio of the height of the human head and that of body posture is ought to be evaluated. The distinguished four types of postures according to the ratio were standing, sitting or kneeling, sitting cross-legged, and other postures. Second, sitting and kneeling were judged according to the 3D spatial relation of the special points. Finally, feature vectors were extracted, transformed, and input to the BPNN according to the characteristics of the other postures, and bending and lying were recognized. Experiments proved the timeliness and robustness of the hybrid approach. The recognition accuracy was high, in which the average value was 99.09%. |
Full text from SpringerLink