Deep Learning Models for Skeleton-Based Action Recognition for UAVs

Human action recognition (HAR) is an important task for UAVs for instant decision-making from captured videos. HAR for UAVs is a challenging task due to the UAVs’ motion, attitudes, and view changes during flight. Moreover, UAVs’ video sequences may suffer from blurs and low resolution. All these issues cause difficulty in HAR for UAVs, necessitating the quest for the HAR method that considers UAV data characteristics. In this paper, we revisit some state-of-the-art deep learning methods and evaluate their performance on the UAV-Human dataset- the largest public UAV dataset up to now. Based on the evaluation, we propose a new framework that combines AAGCN and MS-G3D through a Feature Fusion module for data pre-processing in all streams. Experimental results show that our proposed method outperforms state-of-the-art methods on the UAV-Human dataset.