|
|
Machine Learning-Based Multimodal Data Processing and Mapping in Robotics
Ligocki, Adam ; Duchoň,, František (oponent) ; Saska,, Martin (oponent) ; Žalud, Luděk (vedoucí práce)
This dissertation deals with the application of object detection neural networks on multimodal data in robotics. It aims at three topics: dataset-making, multimodal data processing, and neural network training. The most important is a proposed method that allows creating a large training dataset without an expensive and time-demanding human annotation. The method uses the neural network model trained on the RGB image data and uses multiple sensors' data to create the surrounding map and transfers the annotations of objects detected in the RGB image to the other data domain, like thermal images or point cloud data. Applying this approach, the author generated the thermal image dataset, which contained hundreds of thousands of annotated images, and used them to train the network that outperformed other models trained on human-annotated data. Moreover, the thesis also studies the robustness of object detection in various data domains during difficult weather conditions. The thesis also describes the entire multimodal data processing pipeline that the author created during his Ph.D. studies. That includes developing a unique sensory framework that employs a wide range of commonly used sensors in robotics and self-driving cars. Next, it describes the process of using the sensory framework to make a large-scale publically available open-source navigation and mapping dataset called Brno Urban Dataset. Finally, it covers the description of the custom-made software tools, the Atlas Fusion and the Robotic Template Libarary that the author used to manipulate the multimodal data.
|
|
|
Machine Learning-Based Multimodal Data Processing and Mapping in Robotics
Ligocki, Adam ; Duchoň,, František (oponent) ; Saska,, Martin (oponent) ; Žalud, Luděk (vedoucí práce)
This dissertation deals with the application of object detection neural networks on multimodal data in robotics. It aims at three topics: dataset-making, multimodal data processing, and neural network training. The most important is a proposed method that allows creating a large training dataset without an expensive and time-demanding human annotation. The method uses the neural network model trained on the RGB image data and uses multiple sensors' data to create the surrounding map and transfers the annotations of objects detected in the RGB image to the other data domain, like thermal images or point cloud data. Applying this approach, the author generated the thermal image dataset, which contained hundreds of thousands of annotated images, and used them to train the network that outperformed other models trained on human-annotated data. Moreover, the thesis also studies the robustness of object detection in various data domains during difficult weather conditions. The thesis also describes the entire multimodal data processing pipeline that the author created during his Ph.D. studies. That includes developing a unique sensory framework that employs a wide range of commonly used sensors in robotics and self-driving cars. Next, it describes the process of using the sensory framework to make a large-scale publically available open-source navigation and mapping dataset called Brno Urban Dataset. Finally, it covers the description of the custom-made software tools, the Atlas Fusion and the Robotic Template Libarary that the author used to manipulate the multimodal data.
|
host ::
RSS.