Unmanned aerial vehicles for human detection and recognition using neural-network model

Abbas, Yawar; Al Mudawi, Naif; Alabdullah, Bayan; Sadiq, Touseef; Algarni, Asaad; Rahman, Hameedur; Jalal, Ahmad · 2024 · Crossref

DOI: 10.3389/fnbot.2024.1443678

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of recognizing human actions from unmanned aerial vehicle (UAV)-captured RGB videos, a task complicated by dynamic backgrounds, motion blur, occlusions, and varying camera angles. The authors aim to develop a robust system for applications such as surveillance, sports analysis, and human-robot interaction that does not rely on depth information, which is often unavailable or unreliable in real-world aerial footage. To overcome these limitations, the study proposes a neural-network-based approach that processes aerial RGB data directly, focusing on efficient feature extraction and classification. The methodology involves a multi-stage pipeline beginning with video segmentation into individual frames. These frames undergo preprocessing, including Gaussian blur to reduce noise and grayscale conversion to optimize computational costs and enhance foreground visibility. Human detection is performed using the YOLOv9 algorithm, which identifies human bodies and generates bounding boxes. From the detected silhouettes, the system extracts a human skeleton comprising 15 key points, including the head, neck, shoulders, elbows, wrists, hips, knees, ankles, and belly button. Since standard pose estimators do not detect all landmarks, midpoints are calculated for missing points like the neck. The system then derives specific features from these keypoints, including relative joint angles, distances between joints, 3D point clouds, and fiducial points. These features are optimized using Kernel Discriminant Analysis (KDA) and subsequently classified using a Convolutional Neural Network (CNN). The proposed model was validated on three benchmark datasets: UAV-Human, UCF, and Drone-Action. The system achieved action recognition accuracies of 0.68, 0.75, and 0.83, respectively. The study also compared the performance of various YOLO versions (v1 through v9) for human detection, demonstrating a consistent improvement in accuracy with newer iterations, with YOLOv9 yielding the highest detection precision. The integration of YOLOv9 for detection and CNN for classification proved superior to previous methods that relied on depth data or traditional machine learning classifiers like SVM and Random Forest. The significance of this work lies in its ability to perform accurate human action recognition using only RGB aerial imagery, eliminating the dependency on depth sensors. By leveraging deep learning for both detection and classification, along with KDA for feature optimization, the system offers a more robust and generalizable solution for UAV-based surveillance and analysis. The results demonstrate that processing spatial and temporal relationships through skeletal keypoints and deep neural networks can effectively handle the complexities of drone-captured video, providing a viable framework for real-time applications in security, rehabilitation, and sports analytics.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms