Abstract

Abstract Nowadays Human activity recognition takes a fascinating part in miscellaneous fields of computer vision like medical care, video surveillance, human-computer interface. Recently optical flow were shown to be an efficient feature for action recognition and attained state-of-the-art accuracy on different datasets. In this work, we have taken a deeper look at the combination of action recognition and optical flow. This paper proposes a novel human activity recognition technique based on the 3D dense optical flow from video sequences. Transfer learning and 3D dense optical flow was used in a two-stream neural network architecture. For transfer learning, ResNet152 pre-trained architecture was used. ResNet152 extracted several fine grained features from the dense optical flow. The proposed method was tested with UCF-101 dataset and outperformed the existing state-of-the art methods in terms of accuracy.KeywordsOptical flowHuman activity recognitionResidual networkTransfer learning

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call