https://libeldoc.bsuir.by/handle/123456789/45774| Title: | UNetX: Real-time Pedestrian Crosswalk Segmentation on Mobile Device |
| Authors: | Adaska, E. Lechanka, A. |
| Keywords: | материалы конференций;conference proceedings;autonomous car;pedestrian crosswalk;image segmentation;deep neural network;U-Net;depthwise separable convolution |
| Issue Date: | 2021 |
| Publisher: | UIIP NASB |
| Citation: | Adaska, E. UNetX: Real-time Pedestrian Crosswalk Segmentation on Mobile Device / Adaska E., Lechanka A. // Pattern Recognition and Information Processing (PRIP'2021) = Распознавание образов и обработка информации (2021) : Proceedings of the 15th International Conference, 21–24 Sept. 2021, Minsk, Belarus / United Institute of Informatics Problems of the National Academy of Sciences of Belarus. – Minsk, 2021. – P. 75–78. |
| Abstract: | This paper presents a lightweight deep neural network that segments pedestrian crosswalks on an image in realtime. It is based on U-Net architecture with all its convolution layers substituted with depthwise separable convolution ones. This neural network was trained and tested against a set of manually segmented 3083 road images — with and without crosswalks. The resulting network has only 383K parameters and runs at 35 FPS on a mobile phone. The Jaccard index (IoUmetric) on the validation set is 0.9138. |
| URI: | https://libeldoc.bsuir.by/handle/123456789/45774 |
| Appears in Collections: | Pattern Recognition and Information Processing (PRIP'2021) = Распознавание образов и обработка информации (2021) |
| File | Description | Size | Format | |
|---|---|---|---|---|
| Adaska_UNetX.pdf | 1.68 MB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.