Multi-Task Learning for iRAP Attribute Classification and Road Safety Assessment (CROSBI ID 700269)
Prilog sa skupa u zborniku | izvorni znanstveni rad | međunarodna recenzija
Podaci o odgovornosti
Kačan, Marin ; Oršić, Marin ; Šegvić, Siniša ; Ševrović, Marko
engleski
Multi-Task Learning for iRAP Attribute Classification and Road Safety Assessment
We address the automatic recognition of road safety attributes according to the iRAP methodology. We formulate the problem as a separate multi-class classification of each iRAP attribute in georeferenced video clips that correspond to particular road segments. We propose a solution based on an efficient multi-task model with shared features, which can recognize all attributes with a single forward pass and learn in an end-to-end fashion. We perform experiments on a novel real dataset acquired along 1850 km of public roads in Bosnia and Herzegovina, in which all iRAP attributes have been annotated by human experts. We express recognition accuracy as per-attribute macro-F1 scores due to a significant class imbalance present within most attributes. We thoroughly validate different variants of our model, analyze the contributions of several hyper-parameters, and report recognition accuracy on the independent test set.
Computer vision, road safety assessment, video classification
nije evidentirano
nije evidentirano
nije evidentirano
nije evidentirano
nije evidentirano
nije evidentirano
Podaci o prilogu
9294305
2020.
objavljeno
10.1109/ITSC45102.2020.9294305
Podaci o matičnoj publikaciji
Proceedings of the 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC 2020)
Lu, Meng ; Wang, Yibing ; Barth, Matthew
Piscataway (NJ): Institute of Electrical and Electronics Engineers (IEEE)
978-1-7281-4149-7
2153-0009
Podaci o skupu
23rd IEEE International Conference on Intelligent Transportation Systems (ITSC 2020)
predavanje
20.09.2020-23.09.2020
Rodos, Grčka
Povezanost rada
Računarstvo