search

erdos-project / pylot

Modular autonomous driving platform running on the CARLA simulator and real-world vehicles.
https://pylot.readthedocs.io/
Apache License 2.0
451 stars 126 forks source link

Improve lane detection #66

Open ICGog opened 4 years ago

ICGog commented 4 years ago

We need lane detection for when we don't have access to a fully detailed HD Map. To address this limitation, we can either extract lanes from segmented frames, or integrate Lanenet into Pylot (or any better alternative). We've previously tested Lanenet with this script, but we haven't wrapped it in an operator.

sukritkalra commented 4 years ago

I can take a look at how good LaneNet works with real data from Vijay’s camera frames. Is there any other reason why we would need complete semantic segmentation?

ICGog commented 4 years ago

I don't really see any other reason at the moment.

ICGog commented 4 years ago

Lanenet works pretty well, but its postprocessing code is slow (around 420 ms). In particular, https://github.com/ICGog/lanenet-lane-detection/blob/master/lanenet_model/lanenet_postprocess.py#L163 takes around 350 ms.

We'll have to investigate this if we end up using LaneNet heavily.