Recognizing human actions using a new descriptor based on spatial–temporal interest points and weighted-output classifier

The bag of interest points (BIPs) approach is a good strategy for human action recognition, but it ignores much information contained in the spatial–temporal interest points (STIPs), while the lost information is helpful for classification. In this paper, a new action descriptor based on the STIPs is proposed: histogram of interest point locations (HIPLs). HIPL reorganizes STIPs and reflects the spatial location information, and can be viewed as a useful supplement to the BIP feature. Multiple features including BIP and HIPL are extracted to describe human actions, however, it leads to over-fitting easily by combining them directly because the dimension of feature vector is too high. To overcome this problem, a novel classifier combination framework is developed to integrate the multiple features, and AdaBoost and sparse representation (SR) are used as basic algorithms. Experiments on KTH and UCF sports datasets which are two benchmarks in human action recognition, show that our results are either comparable to, or significantly better than previously published results on these benchmarks.