|
|
Vehicle Make and Model Recognition in Image
Hrivňák, Marek ; Hradiš, Michal (referee) ; Herout, Adam (advisor)
This thesis focuses on training convolutional neural network for vehicle recognition in image, preparation of training data and improvement of classification accuracy. Solution focuses on effect of using 2D bounding box and data augmentation for better recognition accuracy. In this thesis, I also elaborate the comparison with papers using 3D bounding box and showing, my method approaches in some cases even outperforms method using 3D bounding box. BoxCars116k data set is used, which is freely available and collected by the GRAPH@FIT research group. In order to support the main data set, I also collected some vehicle images. As a result of the analysis, it is observed that accuracy of vehicle recognition increased 8% points in comparison with other convolutional neural networks without the proposed modifications. As part of my thesis I also performed several experiments, which show effect of different factors on classification accuracy.
|
|
|
Back-End Design for Analytical Dashboard of Positioning System
Hrivňák, Marek ; Bardonek, Petr (referee) ; Šimek, Václav (advisor)
This thesis deals with the optimization of the original solution and the design of a new solution for the storage and processing of positional data for the tool Sage Analytics of RTLS developed by Sewio Networks s.r.o. The objective of this study is to find and implement a solution to reduce the production time of Sage Analytics metrics. The optimized original solution provides in a very simple way (without the need to modify the implementation) significant acceleration in the production of metrics (for a time interval of 24 hours of displayed data) on average by up to 503 %. The proposed solution uses the InfluxDB database to store positional data and modifies the data processing in such a way, that it is compatible with the use of a new method of storing and retrieving positional data. The new solution also includes tests to check the correctness of the proposed solution. The application of this solution brings acceleration in the production of metrics (for the time interval of 24 hours of displayed data) from 725 % up to 2085 % and on average up to about 1010 %. Part of the work is also the performance of several experiments, which aim to reveal the reasons for the duration of metrics in Sage Analytics.
|
|
|
Back-End Design for Analytical Dashboard of Positioning System
Hrivňák, Marek ; Bardonek, Petr (referee) ; Šimek, Václav (advisor)
This thesis deals with the optimization of the original solution and the design of a new solution for the storage and processing of positional data for the tool Sage Analytics of RTLS developed by Sewio Networks s.r.o. The objective of this study is to find and implement a solution to reduce the production time of Sage Analytics metrics. The optimized original solution provides in a very simple way (without the need to modify the implementation) significant acceleration in the production of metrics (for a time interval of 24 hours of displayed data) on average by up to 503 %. The proposed solution uses the InfluxDB database to store positional data and modifies the data processing in such a way, that it is compatible with the use of a new method of storing and retrieving positional data. The new solution also includes tests to check the correctness of the proposed solution. The application of this solution brings acceleration in the production of metrics (for the time interval of 24 hours of displayed data) from 725 % up to 2085 % and on average up to about 1010 %. Part of the work is also the performance of several experiments, which aim to reveal the reasons for the duration of metrics in Sage Analytics.
|
|
|
Vehicle Make and Model Recognition in Image
Hrivňák, Marek ; Hradiš, Michal (referee) ; Herout, Adam (advisor)
This thesis focuses on training convolutional neural network for vehicle recognition in image, preparation of training data and improvement of classification accuracy. Solution focuses on effect of using 2D bounding box and data augmentation for better recognition accuracy. In this thesis, I also elaborate the comparison with papers using 3D bounding box and showing, my method approaches in some cases even outperforms method using 3D bounding box. BoxCars116k data set is used, which is freely available and collected by the GRAPH@FIT research group. In order to support the main data set, I also collected some vehicle images. As a result of the analysis, it is observed that accuracy of vehicle recognition increased 8% points in comparison with other convolutional neural networks without the proposed modifications. As part of my thesis I also performed several experiments, which show effect of different factors on classification accuracy.
|
guest ::
