University of Houston

Recent research in human action recognition is focused more on local motion descriptor based approaches, wherein, a number of feature points are found across the human body and their local motion is used as the underlying model for action recognition. This makes recognition less dependent on surrounding environment (illumination, camera position etc) and hardware.

The goal of this thesis is to study different local - motion based approaches for human action recognition. For the purpose of this study, we have build a common kNN classifier based evaluation framework which evaluates a number of combinations of different spatio-temporal interest point detectors and descriptors for human action recognition.

A descriptor is used to model the local motion of feature points distributed across the body where the action is occurring. We have evaluated three different detectors and five different descriptors using a common k Nearest Neighbor (kNN) classifier based evaluation framework.The model is then trained on a dataset comprising of a number of instances of each action to be recognized and tested using kNN classifier, which classifies a test video into one of the actions for which the model has been trained. Using this framework, we study the effectiveness of different descriptors and detectors on three different datasets: KTH Actions, UCF Sports Actions and UT-Tower Dataset, and provide results for each.