Author Country (or Countries)



Despite the vast amount of research on the analysis of existing and ongoing human activity, there are still significant challenges worthy of address. In this paper, an innovative approach for human action recognition based on discriminative models like CRFs, HCRFs and LDCRFs is proposed. To handle human action recognition, different number of window size ranging from 0 to 7 are applied using a compact computationally-efficient descriptor as statistical chord-length features (SCLF), in addition to optical flow motion features that derived from 3D spatio-temporal action volume. Our experiment on a standard benchmark action KTH, as well as our IIKT dataset show that the recognition rate, and the reliability of human activity is improved initially as the window size increase, but degrades as the window size increase further. Furthermore, LDCRFs is robust and efficient than CRFs and HCRFs, in addition to problematic phenomena than those previously reported. It also can carry out without sacrificing real-time performance for a wide range of practical action applications.

Suggested Reviewers