In order to overcome the limitation of Bag of Features (BoF),which ignores the space time relationship of local features in human action recognition,a space time coding (STC) method for local feature was proposed by involving the space time locations of local features into feature coding phase to directly model their space time relationship.First,the local features were projected into a sub space-time-volume (sub-STV) to obtain their space time coordinates.Second,their appearance information and space time locations were encoded simultaneously.After that,the statistics results generated by feature pooling upon these codes were utilized for action classification.To achieve better performance,the multi-scale STC and locality-constrained STC were also proposed.In action classification,a locality-constrained block sparse representation classifier (LBSRC) was adopted to improve the action recognition accuracy.The experimental results on KTH,Weizmann,and UCF sports benchmark datasets showed that the proposed methods can effectively represent the space time relationship of local features and improve the action recognition accuracy.
