A Bird Surveillance Approach Based on Deep Learning

Résumé

Monitoring wildlife, particularly migratory birds in wetlands and their natural habitat, presents significant challenges due to the unique requirement for real-time, high-fidelity data collection in remote, isolated and harsh environments with limited energy and communication resources. This thesis proposes a deep learning and WSN-based, efficient approach for migratory birds’ surveillance in wetland ecosystems. The system incorporates advanced Region of Interest (ROI)-based detection, adaptive video compression, and long-range wireless communication “LoRaWAN”, to optimize resource usage without sacrificing data fidelity. In addition, it is leveraging two artificial intelligence algorithms “You Only Look Once (YOLOv8)” and “Learning to Count Everything” (LTCE), each with specialized task, for detection and counting. The proposed work employs at the sensor level a multi-level ROI-based compression technique to prioritize the most significant regions within a video frame to reduce bandwidth usage and power consumption while maintaining high fidelity of important data. Followed by a dynamic video encoding approach based on the extracted regions previously to compress the extracted regions with different quality factors based on their importance, this maintains high quality within the important regions and reduce the data usage on non-important regions. A LoRaWAN communication module will transmits the data to the base station over long-range, in real-time using less energy. At the base station, the system employs two image processing algorithms for object detection and counting, YOLOv8 is for real-time species identification and a “Single Shot” “LTCE” for objects counting. To mitigate potential overfitting and ensure effective generalization across diverse environmental conditions and datasets, we choose to use lightweight models and train them for efficiency across diverse and general datasets that provides a wide range of species, and natural environment conditions so that the models are more adaptive to multiple categories and situations. Finally, we create a loss model to evaluate our models across varying environmental factors (such as distance, obstacles, humidity level, and vegetation density). Comprehensive experiments were conducted to assess the system’s performance across various video sequences and environmental conditions. Results indicate that the proposed work outperforms conventional video encoding techniques in terms of energy efficiency, data compression, and transmission latency, while maintaining high-quality image reconstruction. Furthermore, statistical evaluations using PSNR, SSIM, and BRISQUE metrics confirm the system’s effectiveness in preserving critical visual information. The performance of the chosen AI algorithms is evaluated using relevant metrics. By bridging the gap between smart surveillance and ecological monitoring, this research contributes to the development of scalable, energy-efficient wildlife monitoring solutions. The findings highlight the potential for deploying AI-driven, resource-aware systems in conservation efforts, aiding researchers in better understanding migratory bird behaviors and environmental impacts.

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 3

Téléchargements : 0