• accelerator
• AI
• computer vision
• dataset
• depth
• disparity
• embeded systems
• neural networks
• OAK-D
• raspberry pi
• wildlife classification
• Yolo

Understanding the animals’ behavioral responses to environmental changes is a frequent
goal in animal ecology and has important implications for conservation. Also the increased rate of encountering of human-populated zones with wild animals have made it very important to monitor how animals are responding to these changes and to search for new strategies to manage the situation.
Only in the Tuscany region in the last years with the decline of rural cultivation animals such as wild boars and deers are much more present in human-areas leading to important challenges like crop destruction, road accidents and the risk of disease transmission.
In this context trying to combine the data with the advancements in technology can lead us to important studies. The latest advancements in hardware devices like drones,
smartphones and camera traps have made possible the generation of large amounts of data.
When combined with developments in Deep Learning, this data
improves the effectiveness of machine learning models. However, traditional Artificial Intelligence and Machine learning solutions face challenges because of high latency, power consumption, data centralization and privacy concerns. To address these issues there are many research projects aiming to create a distributed edge computing infrastructure for AI applications in rural and park areas. This thesis focuses on two main issues : (I) the factors that lead to a successful construction and structuring of high quality datasets, explicitly designed and created for classifying animal species through images captured by camera traps (ii) doing some simulations using these datasets on a camera trap with Raspberry Pi 4 and an OAK-D camera which its composed by a stereoscopic camera and an Intel Movidius Myriad accelerator; to see some behavioral aspects of wild life and also to try to integrate some distance estimation data than can be further used to extract useful information like size, age or sex of the animal, depending on the use case. I evidence that an effective and quality dataset should have more than just image collection; it should integrate detailed and
explicit metadata related to capture conditions such as date, geographical location, time and camera specifications. These metadata apart from improving the model accuracy also
reveal unexpected insights like correlations with environmental factors. My experiments
uncovered interesting characteristics about the animal that was part of the original
objective, underscoring the significance of a well-structured dataset in wildlife monitoring and management and also the importance of image size in the pre-processing dataset in order to have some practical performance on the Edge.