Abstract | ||
---|---|---|
We present a novel method for outdoor monocular visual mapping and localization, by detecting and probabilistically parameterizing road lanes. To present road lane information, we use cascaded deep models for detection on keyframes, and rely on piecewise cubic Catmull-Rom splines for parameterization in complicated environments. Additionally, we propose a maximum-a-posteriori estimation framework to conduct offline mapping, by jointly optimizing the lanes’ geometric appearance and vehicle poses. The computed maps can be used to perform semantic localization, without relying on traditional visual point features. The underlying key design motivation is that, compared to visual point features, road semantic information is naturally highly compact and of long-term consistency (in terms of existence, appearance, and so on). Results from both simulated and real-world experiments show that our method is able to enhance estimation accuracy with cost-effective maps by wide margins. |
Year | DOI | Venue |
---|---|---|
2020 | 10.1007/978-3-030-71151-1_33 | ISER |
Keywords | DocType | Citations |
Semantic mapping,Piecewise,Computer vision,Cubic Hermite spline,Point (geometry),Consistency (database systems),Monocular,Parameterized complexity,Computer science,Key (cryptography),Artificial intelligence | Conference | 0 |
PageRank | References | Authors |
0.34 | 0 | 3 |
Name | Order | Citations | PageRank |
---|---|---|---|
Sheng Yang | 1 | 188 | 15.53 |
Yiming Chen | 2 | 187 | 22.75 |
Mingyang Li | 3 | 270 | 17.60 |