Recognizing human actions based on gist descriptor and word phrase

The precision of human action recognition mainly depends on the discrimination of action feature descriptors. The descriptors which contained both global and local information usually can be competent for classifying actions. A novel descriptor based on Gist descriptor and word phrase is proposed. Gist descriptor captures the structure information of action and its vectors imply the location relationship of local grids. According to the location which the Gist vectors attribute to, the features are divided into four segments, and word phrases are respectively constructed using bag of words method. By accumulating all the phrases, the final action descriptor is produced. Through SVM classifier our method obtains impressive results on KTH dataset.