Action Recognition Using Spatio-Temoral Descriptors

Action recognition is one of the most challenging problems in computer vision. Action recognition has wide applications in domains such as robotics, video retrieval, visual surveillance and human-computer interaction. The recognition methods have been developed into probability models based on spatio-temporal descriptors from body models and image models.Experiments show that probability models based on spatio-temporal descriptors have a higher recognition accuracy, and we studied spatio-temporal descriptors in this paper. A topic model which was used in document retrieval was also proposed for action recognition. By searching the key points in videos, we could get video patches and generate local spatio-temporal descriptors. According to the relationship between descriptors, we could get several topics and finally label an action by making use of these topics.Harris3D was used to searching key feature points which have a relatively high gradient values in videos. We created three kinds of descriptors using these video patches:angle interval descriptors, three-dimension projection descriptors and compound descriptors. During training and test stages, we adopted bag of words model and topic model.We tested the above three kinds of descriptors and two models on two public action databases with two classifiers--naive bayes and support vector machine--and compared them with each other. Experiments show that each descriptors has both advantages and disadvantages and topic model has a relatively higher recognition accuracy than the bag of words model.