Human activity recognition using overlapping multi-feature descriptor

An efficient overlapping multi-feature descriptor and classification scheme for human activity recognition is introduced. The descriptor is constructed by overlapping global feature combinations of multiframes using a Hankel matrix representation. The descriptor captures the local and temporal information while overcoming the limitations of global features using an overlapping combination scheme. In addition, a random forests classifier is used to cope with noise in the descriptor that can be obtained from no-activity frames in a video. Using this framework, it is shown that the approach outperforms the state-of-the-art methods using the KTH dataset and a much more complex human interaction dataset.