Distributed Implementation and Performance Assessment of a Deep Neural Network Inference System for Wildlife Monitoring

This thesis explores the application of distributed inference in edge computing technologies to enhance wildlife monitoring systems, with a particular focus on the YOLOv8 object detection model. In order to understand the application of distributed machine learning in edge computing, the thesis explores the concept of distributed inference. This approach involves processing data closer to the data source, desirably on the on-field devices, in this case the NVIDIA Jetson Orin Nano Super Developer Kit. Edge devices enables immediate data analysis of what is captured by cameras and sensors, reducing the need for constant communication with a distant data center. At the core of this research is the YOLOv8 (You Only Look Once, version 8) object detection model that excels in real-time object detection tasks. The YOLOv8 paradigm has been chosen because it was provided by the researchers as pre-trained model, that have been trained on the specific wild dataset. Whenever the battery on one device runs low, its workload can be offloaded to other devices, extending its operational time. By distributing the model, we can balance power consumption across devices and also monitoring factors such as latency, network bandwidth and streaming performance. There are the three key points where the architecture has been divided into two: the backbone, the neck, and the head. The assessment of those metrics are crucial factors in the practical deployment of distributed machine learning systems in remote or rural areas.