Proiect finantat de Bosch.
Director proiect: prof. dr. Laura Diosan.
Perioada de derulare: octombrie 2021 – septembrie 2022
Abstract:
Switch identification is the task of detecting in images the location of switches on train tracks and classifying the direction in which they are oriented. Some attempts to solve this task have been made using rudimentary computer-vision approaches instead of using the power of deep neural network architectures. Other approaches that use deep networks fail to provide a complete system for identifying switches. This project consists of the proposal of a system for detecting switches from the ego-view of the train using a two-step process by dividing the problem into a segmentation task and a classification task. The obtained system is called SWIDER. We also propose three hyper-parameters which filter and enhance the most comprehensive dataset from the rail scene, namely RailSem19, and show that this improvement leads to better results for the switch classification task. Our results show a considerable increase in the performance of rails segmentation and switch classification when compared to different works from the same area of study. There is room for improvement for the switch detector.
The results of this work were presented, disseminated and awarded in various events such as the scientific public speaking contest Games of Science 2022 organized by The British Council Romania, the Students Scientific Communication Session 2022 (SSCS) where it received the third prize, and the Eastern European Machine Learning Summer School 2022 sponsored by companies such as DeepMind, Google or Vinted.