Integrated Detection for Behavior Recognition in Videos

Abstract

In computer vision, activity recognition in video has attracted researchers due to its variety of applications such as, human computer interaction, video retrieval and surveillance system. At the same time it’s  really challenging to detect actions in real-time world, due to its complex motion style and background litters. This causes several confusion. Videos of high-dimensionality also limit the performance of recognition. Many number of features are required in order to obtain good action representation and also to minimize asynchrony amongst stream data. The feature of Motion Heat Map (MHM)is  considered to represent group activity and implement motion information to signify the trajectory. Global motion  pattern information is obtained using optical flow. The gained feature vectors from optical flow and MHM are combined and fed as input to bag of words approach in order to detect normal and abnormal frames. For simulation and validation, widely used dataset like UMN is considered to validate the proposed model.