Activity Recognition Using Deep Recurrent Neural Network on Translation and Scale-Invariant Features
Md. Zia Uddin, Weria Khaksar, Jim Tørresen · 2018
Recent advances in image processing and computer vision have driven to numerous initiatives to recognize human activities from video data. This work proposes a human activity recognition approach using robust translation and scale-invariant body silhouette features recurrent neural network. First, Human body silhouette is extracted from a depth image after background subtraction. Then, body parts are segmented using random forests to get corresponding body skeleton in the image. Furthermore, scale-invariant skeleton features are extracted by representing the body joints in the spherical coordinate system. Then, the skeleton features are augmented with the motion features of the skeleton in consecutive frames. To combine with the skeleton features, Radon transformation is applied on the depth silhouettes to extract translation and scale-invariant silhouette features. The robust features extracted from the depth image sequences are then applied to a deep recurrent neural network for activity training and recognition. The proposed approach shows the superiority over other approaches by achieving greater than 98 % mean recognition rate on private and public datasets where others can yield around 95%.