Earticle

초록

영어

The aerial view diverse action recognition (AR) benchmark provides a valuable resource for researchers and developers in computer vision (CV) for human actions recognition (HAR) from an aerial perspective. With the increasing use of unmanned aerial vehicles (UAVs) for surveillance, delivery, search, and rescue, a robust understanding of human actions from an aerial view is crucial. Existing datasets lack representation of common outdoor actions and are unsuitable for intelligent UAVs. This article proposes a dataset that captured various actions from diverse viewpoints and in different environments. The dataset includes three viewpoints (Top, left, and right) allowing angle-invariant algorithm development. State-of-the-art algorithms (3D, and 2D convolutions with sequential learning) are evaluated on the dataset. The proposed model demonstrates exceptional performance with high accuracy (87.5%), precision (86.3%), and recall (87.2%) rates. The robustness of the model is showcased through real-time testing, indicating that the proposed dataset and model contribute to advancing research from drone view AR and have the potential to enhance surveillance and other UAV applications.

목차

Abstract
1. Introduction
2. Method
2.1 Dataset Collection and Preprocessing
2.2 Proposed Model
3. Experiment result
3.1 Experimental setting
3.2 Ablation study
4. Conclusions and future work
Acknowledgment
References

저자

• Muhammad Munsif [ Sejong University ]
• Haseeb Ali Khan [ Sejong University ]
• Minje Kim [ Sejong University ]
• Fatema Rahimi [ Sejong University ]
• Sana Parez [ Sejong University ]
• Mi Young Lee [ Sejong University ]
• Soo-Mi Choi [ Sejong University ]
• Jong Weon Lee [ Sejong University ] Corresponding Author

참고문헌