Analysis of vertebrates’ activity by machine learning

Activity Recognition (AR) is nowadays a fervent research area which gives many new challenges to deal with. In particular, we focus our attention on the development and analysis of classifiers based on Machine Learning (ML) approaches in the areas of Human Activity Recognition (HAR) and biologging. The literature already presents many different approaches for ML classifiers for HAR problems, especially in monitoring of humans in domestic environments. Less common is the application of ML classifiers in biologging, which is still commonly studied through traditional methods. In our research we applied classifiers implemented by ML approaches from the classes of Neural Network (NN) and of Support Vector Machine (SVM). The classification of the activities was performed over time−series collected by sensory devices. The devices were worn by each subject of the case studies considered. The classifiers were configured and tuned specifically for each case study at hand. In this thesis we dealt with four cases of study: with humans to identify daily activities, with tortoises to identify the digging activity, and with penguins and seals to identify the prey handling activity. These case studies covered an heterogeneous set of both HAR and biologging problems. The classifier applied via shift-window over the time and specifically tuned by input sequences (windows shifted over the input time−series) is implemented by Input Delay Neural Network (IDNN), Convolutional Neural Network (CNN), and SVM which naturally deal with these input. For the same case studies, we implemented the classifier by models from the Recurrent Neural Networks (RNN) class, which naturally apply over streams by taking advantage from their internal memory. We evaluated each implementation of the classifier by means of its accuracy and F1 score reached in classification, and by assessing its feasibility for its use into embedded devices in terms of memory space. We demonstrated that with sequences as input, the IDNN model provides a good trade off between performance (accuracy of the model) and feasibility (memory footprint of the model). Instead with streams we observed that the Echo State Network (ESN) reaches a good performance and it is feasible as well because the reservoir can be kept small without a significant penalty in terms of performance of classification. The results of this analysis would contribute to improve future activity recognition methods. We showed that it was possible to implement efficient classifiers in selected real−world case studies. In particular, such efficiency of the classifiers allows to meet the performance requirements of real applications enabling the embedding of the classifiers into low−power devices. In perspective, we believe that this research will support future research directions with the focus on stimulating research in the directions of animals’ monitoring/protection.