Integrated neural network framework for multi-object detection and recognition using UAV imagery

Alshehri, Mohammed; Xue, Tingting; Mujtaba, Ghulam; Alqahtani, Yahya; Almujally, Nouf Abdullah; Jalal, Ahmad; Liu, Hui · 2025 · OpenAlex-citations

DOI: 10.3389/fnbot.2025.1643011

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 accurate vehicle analysis from Unmanned Aerial Vehicle (UAV) imagery, which is critical for applications such as intelligent traffic management, urban planning, and autonomous navigation. The authors identify inherent difficulties in UAV-captured video, including small target sizes, occlusions, cluttered backgrounds, motion blur, and fluctuating lighting conditions, which hinder conventional perception systems. To overcome these issues, the study proposes a fully end-to-end deep learning framework specifically optimized for UAV-based traffic monitoring. The primary motivation is to integrate multiple specialized neural network modules into a unified pipeline that handles spatiotemporal dynamics and ensures robust performance in diverse, real-world aerial scenarios. The proposed framework integrates several state-of-the-art models, each tailored to a specific task. RetinexNet is used for preprocessing to normalize lighting and enhance contrast. High-Resolution Network (HRNet) performs semantic segmentation to preserve high-resolution spatial details and separate vehicles from backgrounds. Vehicle detection is handled by YOLOv11, chosen for its speed and precision with small objects. Deep SORT is employed for reliable vehicle tracking, while CSRNet facilitates high-density vehicle counting. Long Short-Term Memory (LSTM) networks predict vehicle trajectories by capturing temporal patterns, and a combination of DenseNet, SuperPoint, and an AutoEncoder extracts robust features. Finally, Vision Transformers (ViTs) perform vehicle classification using attention mechanisms. The system was evaluated on two benchmark datasets, AU-AIR and Roundabout, which are known for their complexity in aerial traffic analysis. The results demonstrate that the integrated framework significantly improves accuracy, reliability, and efficiency compared to previous benchmarks. On the AU-AIR dataset, the system achieved a detection accuracy of 97.8%, a tracking accuracy of 96.5%, and a classification accuracy of 98.4%. On the Roundabout dataset, it reached 96.9% detection accuracy, 94.4% tracking accuracy, and 97.7% classification accuracy. These findings indicate that the modular architecture effectively handles challenging conditions such as occlusion, variable lighting, and scale variations. The seamless integration of spatial and temporal information allows the system to maintain high performance across diverse aerial traffic scenarios. The significance of this work lies in its demonstration that a unified, multi-model deep learning pipeline can effectively address the multifaceted challenges of aerial vehicle perception. By combining specialized architectures for enhancement, segmentation, detection, tracking, counting, trajectory prediction, and classification, the framework offers a scalable and robust solution for real-time deployment. The study highlights the importance of integrating complementary neural models to handle the dynamic and noisy nature of UAV imagery, providing a strong foundation for future developments in intelligent transportation systems, surveillance operations, and autonomous navigation.

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