Publicação
Anomaly detection in fleet service vehicles: improving object segmentation
| Resumo: | The present dissertation is inserted in a BOSCH project in which the global focus is au tonomous driving. The project is divided in multiple phases, being the main focus of this dissertation object detection and segmentation inside fleet service vehicles. The objective is to detect/segment objects and dirt left inside a vehicle, warning the commuter if they forgot an object inside or the administrators if the vehicles need to be cleaned. To train the models, BOSCH provided an initial dataset containing a small set of annotated images. This dataset contains pictures of a vehicles cockpit with many diverse objects and dirt. One of the goals for BOSCH is to increment this dataset with more images. Hence, in this project several state of the art segmentation methods were thoroughly studied and analysed, with two of them being selected for further exploration: DeepExtremeCut and FgSegNet v2. The main objective is to see to what extent can these methods be used in a semi automatic process to segment more images, thereby increasing the initial dataset. DeepExtremeCut works by using a framework in which, after the model is trained, it allows us to click on four extreme points in the desired object, producing the segmentation. This method produced reliable segmentations, however it requires human intervention both for the initial segmentation and verification of the output. Hence, it was not regarded as a good solution for a future augmentation of the BOSCH dataset. Regarding FgSegNet v2, this later method does not require any initial annotation of the input images. Under this approach only a final verification and possible rectification is required. Therefore, this method meets the requirements defined by BOSCH for a dataset expansion solution. An ablation study is also presented for FgSegNet v2, analysing its three stages: (i) Encoder, (ii) Feature Pooling Module and (iii) Decoder. The result of this study is a proposal of a variation of the aforementioned method called Mod FgSegNet. It was also compared with state of the art methods in public datasets. Three datasets are used for testing: CDNet2014, SBI2015 and CityScapes. In CDNet2014 we got an overall improvement when compared to the state of the art, particularly in the LowFrameRate subset. Regarding SBI2015 the overall results are lower in comparison with the top state of art, while in CityScapes some promising results are presented. |
|---|---|
| Autores principais: | Morais, Joel Tomás |
| Assunto: | Segmentation Object detection Bounding box |
| Ano: | 2021 |
| 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: | The present dissertation is inserted in a BOSCH project in which the global focus is au tonomous driving. The project is divided in multiple phases, being the main focus of this dissertation object detection and segmentation inside fleet service vehicles. The objective is to detect/segment objects and dirt left inside a vehicle, warning the commuter if they forgot an object inside or the administrators if the vehicles need to be cleaned. To train the models, BOSCH provided an initial dataset containing a small set of annotated images. This dataset contains pictures of a vehicles cockpit with many diverse objects and dirt. One of the goals for BOSCH is to increment this dataset with more images. Hence, in this project several state of the art segmentation methods were thoroughly studied and analysed, with two of them being selected for further exploration: DeepExtremeCut and FgSegNet v2. The main objective is to see to what extent can these methods be used in a semi automatic process to segment more images, thereby increasing the initial dataset. DeepExtremeCut works by using a framework in which, after the model is trained, it allows us to click on four extreme points in the desired object, producing the segmentation. This method produced reliable segmentations, however it requires human intervention both for the initial segmentation and verification of the output. Hence, it was not regarded as a good solution for a future augmentation of the BOSCH dataset. Regarding FgSegNet v2, this later method does not require any initial annotation of the input images. Under this approach only a final verification and possible rectification is required. Therefore, this method meets the requirements defined by BOSCH for a dataset expansion solution. An ablation study is also presented for FgSegNet v2, analysing its three stages: (i) Encoder, (ii) Feature Pooling Module and (iii) Decoder. The result of this study is a proposal of a variation of the aforementioned method called Mod FgSegNet. It was also compared with state of the art methods in public datasets. Three datasets are used for testing: CDNet2014, SBI2015 and CityScapes. In CDNet2014 we got an overall improvement when compared to the state of the art, particularly in the LowFrameRate subset. Regarding SBI2015 the overall results are lower in comparison with the top state of art, while in CityScapes some promising results are presented. |
|---|