Insights into aerial intelligence: assessing CNN-based algorithms for human action recognition and object detection in diverse environments

• Published in: Multimedia tools and applications Vol. 84; no. 16; pp. 16481 - 16523
• Main Authors: Maheriya, Krunal, Rahevar, Mrugendrasinh, Mewada, Hiren, Parmar, Martin, Patel, Atul
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.05.2025; Springer Nature B.V
• Subjects: Algorithms; Artificial neural networks; Clutter; Comparative studies; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Human activity recognition; Human motion; Machine learning; Military operations; Moving object recognition; Multimedia Information Systems; Occlusion; Special Purpose and Application-Based Systems; Urban planning; Research challenges; Human action recognition; CNN; Surveillance; Object detection; Future directions; Aerial view
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-19611-z

Today’s era follows a data-driven decision process for large-scale environment analysis. Aerial view-based decision process plays a key role in various domains including surveillance, disaster responses, city planning, and military operations. One of the important and challenging data-driven processes is human action recognition and object detection from the aerial view. The major challenge in aerial view is its notable variance in the scale and perspective of humans and objects. The size of an individual depends on the distance from the camera making object location and human action recognition difficult. In addition, low resolution or poor visibility, occlusion clutter background, and motion blur limit the accurate object detection and localization. CNN’s ability to learn hierarchical representation, take advantage of spatial relationships, manage translation invariance, and leverage transfer learning make it a powerful tool among all for HAR and object detection from an aerial view. We studied applications and progress of aerial view-based human activity recognition (HAR) and object detection (OD) techniques. Various challenges in different environments have been addressed for both HAR and OD examining recent datasets and algorithms spanning seven years. Initially, a categorization of various CNN-based algorithms and their strength and weaknesses are presented for both HAR and OD respectively. Later, a comparative analysis using evaluation matrices and research challenges are discussed in the paper. The comparative study reveals that the self-attention mechanism in CNN performed better among all models but more extensive research is necessary to raise performance standards. A promising future trajectory is outlined for both HAR and OD from an aerial view perspective.