Active Learning for Road Lane Landmark Inventory with Random Forest in Highly Uncontrolled LiDAR Intensity Based Image
- Asier Izquierdo 1
- Jose Manuel Lopez-Guede 22
- 1 Airestudio Geoinformation Technologies Scoop, Albert Einstein Kalea, 44, E6, Oficina 8, 01510 Vitoria-Gasteiz
-
2
Universidad del País Vasco/Euskal Herriko Unibertsitatea
info
Universidad del País Vasco/Euskal Herriko Unibertsitatea
Lejona, España
- Álvaro Herrero (coord.)
- Carlos Cambra (coord.)
- Daniel Urda (coord.)
- Javier Sedano (coord.)
- Héctor Quintián (coord.)
- Emilio Corchado (coord.)
Argitaletxea: Springer Suiza
ISBN: 978-3-030-57801-5, 978-3-030-57802-2
Argitalpen urtea: 2021
Orrialdeak: 862-871
Biltzarra: International Conference on Soft Computing Models in Industrial and Environmental Applications SOCO (15. 2020. Burgos)
Mota: Biltzar ekarpena
Laburpena
Road landmark inventory is becoming an important industry for the maintenance of transport infrastructures among others. Several commercial sensors are available witch include LiDAR sensors allowing to capture up to 1.5 million data point per second.We obtain an intensity based image from the LiDAR point cloud intensity. The landmark detection is posed as a two class classification problem that may be solved by some standard approaches, for example, Random Forest (RF). Besides model parameter selection, a central problem is the construction of the labeled dataset due to human labor cost and the highly uncontrolled conditions of the data capture. We propose an open ended Active Learning approach with a human operator in the loop who can start the Active Learning process when detection quality is degraded by the change in data condition in order to achieve adaptation to them. As an additional contribution, we have assessed the ability of Active Learning to overcome the issues raised by highly class imbalanced dataset, reaching a True Pixel Ratio value of 0.98.