Journal Article

Scale and View Invariant Informative Joint Descriptor (SVI2JD) for Human Action Recognition from Skeleton Data

Bhagyaraju, Vipparthi, Potturu, Sudharsana Rao, Rao, Dustakar Surendra

2023

Overview

One of the biggest challenges in the Human Action Recognition is View-point variations as the actions are captured under multiple views in real time. Furthermore, in the Skelton based action representation, for HAR, only few joints are informative and remaining joints constitutes redundancy. To sort out these problems, this paper proposes a new Action descriptor called as Scale and View Invariant Informative Joint Descriptor (SVI2JD). SVI2JD is a combination of three descriptors; they are namely Self-Similarity Joint Descriptor (SSJD), Informative Joint Descriptor (IJD) and Spherical Joint Descriptor (SJD). SSJD concentrates on the view invariance and employs a Self-Similarity Matrix (S3M) which computes pair wise distance between joints in each frame of action sequence. Next, SJD aims at describing the action through restricted movements of joints because they can’t move beyond particular angle and distance from origin of body. IJD removes the redundant joints those have less contribution towards the action. Further, a 2D Convolution Neural Network Model is proposed for feature extraction and classification. Different fusion rules are employed to fuse the individual results. The Effectiveness of proposed model is demonstrated through its simulation on two challenging datasets; NTU RGB+D dataset and Northwestern UCLA dataset.

• Publisher: International Information and Engineering Technology Association (IIETA)
• Subject: Artificial neural networks / Datasets / Dictionaries / Feature extraction / Human activity recognition / Human body / Invariants / Literature reviews / Redundancy / Self-similarity