Object detection at the edge

Abstract

While monitoring rodents in the Arctic Tundra to evaluate if climate changes affect the ecosystem. The camera-traps of the coat project generates image data in large scale each year. To manually examine the data in regards to label- ing is a tedious and time-consuming job, and a more efficient and automated tool for the task is required. In this thesis we presents the architecture, design and implementation of a object classification model deployed on a small embedded computer, to be used on the gathered image data in order to classify and label the animals at the edge. We conduct transfer-learning on the state-of-the-art pre-trained YOLOv4-tiny model by introducing a labeled COAT image set. We utilize the Convolutional Neural Network of the model to do predictions on a test image set in order to evaluate the model. The result is an application with an embedded model able to predict labels with an accuracy of 96.07% and inference time that classifies it to do so in real-time.