Publicação
Sensor fusion for detection and classification of vehicle impacts
| Resumo: | This thesis was developed as part of a curricular internship at Bosch Car Multimédia SA, in collaboration with the University of Minho, More specifically, an exploratory research thesis aligned with an R&D project that is being developed internally and whose objective is to detect impacts on vehicles that cause damage based on data obtained through sensors, The usefulness of the work developed in this thesis and the project in which it is inserted, in a real context, would be to help vehicle rental companies and car-sharing services to better monitor the conditions of vehicles in their fleets, This would be achieved by placing a device in vehicles that continuously monitored their status, reducing the need for validation and human interaction after use, The main focus of this thesis was to explore how the fusion of information from different sensors could improve the decision-rnaking capabilities of a system whose purpose is to determine whether impacts on the exterior of a vehicle, captured with a set of sensors, resulted in damage, This conjugation of sensory information is known as sensor fusion. ft is a process of combining information from different homogeneous and heterogeneous sensors to obtain a better representation of what is being observed, The approach chosen to achieve this goal consisted of training a set of Machine Learning (ML) algorithms with two distinct datasets, one based only on one data source and the other multiple sources combined. Each pair of models was further evaluated on unseen data, and their performances were compared based on the va lues obtained, Based on the results obtained, it can be said that the application of sensor fusion allowed for better learning by the models, which led to greater robustness in data never seen before. Of the four chosen algorithms, XGBoost, Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN), all had at least one of the evaluation metrics, the Matthews Correlation Coefficient (MCC) and number of False Positive (FP)s in the test set, superior in model-based fused data. Of these, XGBoost and ANN stand out where the results were significantly better in both metrics, |
|---|---|
| Autores principais: | Santos, João Gabriel Lopes dos |
| Assunto: | Sensor fusion Impact detection Machine learning Signal processing Fusão sensorial Deteção de impactos Processamento de sinal |
| Ano: | 2022 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | This thesis was developed as part of a curricular internship at Bosch Car Multimédia SA, in collaboration with the University of Minho, More specifically, an exploratory research thesis aligned with an R&D project that is being developed internally and whose objective is to detect impacts on vehicles that cause damage based on data obtained through sensors, The usefulness of the work developed in this thesis and the project in which it is inserted, in a real context, would be to help vehicle rental companies and car-sharing services to better monitor the conditions of vehicles in their fleets, This would be achieved by placing a device in vehicles that continuously monitored their status, reducing the need for validation and human interaction after use, The main focus of this thesis was to explore how the fusion of information from different sensors could improve the decision-rnaking capabilities of a system whose purpose is to determine whether impacts on the exterior of a vehicle, captured with a set of sensors, resulted in damage, This conjugation of sensory information is known as sensor fusion. ft is a process of combining information from different homogeneous and heterogeneous sensors to obtain a better representation of what is being observed, The approach chosen to achieve this goal consisted of training a set of Machine Learning (ML) algorithms with two distinct datasets, one based only on one data source and the other multiple sources combined. Each pair of models was further evaluated on unseen data, and their performances were compared based on the va lues obtained, Based on the results obtained, it can be said that the application of sensor fusion allowed for better learning by the models, which led to greater robustness in data never seen before. Of the four chosen algorithms, XGBoost, Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN), all had at least one of the evaluation metrics, the Matthews Correlation Coefficient (MCC) and number of False Positive (FP)s in the test set, superior in model-based fused data. Of these, XGBoost and ANN stand out where the results were significantly better in both metrics, |
|---|