ml_laneseg

How to run:

git clone https://github.com/cornellev/ml_laneseg.git

PIXI INSTRUCTIONS: (all from the mllaneseg folder itself) pixi install pixi run build pixi run run-node pixi run start-foxglove

DOCKER INSTRUCTIONS: docker build -t lane_segr_final -f Dockerfile_cuda.jetson .

docker run -it --rm
--runtime nvidia
--privileged
--network host
-v /dev:/dev
-v $(pwd):/ros2_ws
-v ~/zed_models:/usr/local/zed/resources
lane_segr_final bash

colcon build --packages-select zed_ml_inference --symlink-install source install/setup.bash

ros2 run zed_ml_inference lane_segmentation_node3dlanes

FOR TESTING THE MODEL: go to src/zed_ml_inference/zed_ml_inference change code for the picture that you want to test on jsut lfd roadseg byt iself: python3 test_model.py

for the upgraded version with multilane + intersection capabilities: python3 test_model_multilane.py

TOPICS (lane_segmentation_node_final — the pixi run run-node node)

Topic	Type	Notes
lane/camera_raw	sensor_msgs/Image	raw ZED left image
lane/overlay	sensor_msgs/Image	colour overlay + HUD
lane/points/lanes	sensor_msgs/PointCloud2	x,y,z,crossable,point_type,lane_id
lane/costmap	nav_msgs/OccupancyGrid	planner-ready, lane rule applied
lane/costmap/lane_layer	nav_msgs/OccupancyGrid	markings only, rule NOT applied
lane/costmap/lanes_as_obstacles	std_msgs/Bool	current toggle state (latched)
lane/costmap/set_lanes_as_obstacles	std_msgs/Bool	toggle command in

Costmap and the lane toggle

lane/costmap values: -1 unknown, 0 free drivable surface, 100 lethal, anything in between is a soft cost. The grid lies on the ground plane in the camera frame; the ground height is re-estimated from the road surface every frame, so no hand-measured camera height is needed.

lanes_as_obstacles (default true) governs crossable (dashed) markings only. Solid lines stay lethal either way, because crossing them is illegal rather than merely undesirable.

# permit lane changes — dashed lines drop to a soft cost
ros2 topic pub --once /lane/costmap/set_lanes_as_obstacles std_msgs/Bool '{data: false}'

# back to lane keeping
ros2 topic pub --once /lane/costmap/set_lanes_as_obstacles std_msgs/Bool '{data: true}'

# same switch as a parameter
ros2 param set /lane_segmentation_node lanes_as_obstacles false

lane/costmap/lane_layer publishes the markings with the rule not applied (-1 none, 0 crossable, 1 centreline, 100 solid) so a future lane-change planner can make the decision itself instead of relying on the toggle.

Costmap geometry params (set at launch): costmap_resolution (0.10 m), costmap_forward_min/max (-2/20 m), costmap_lateral_half (6 m). Live-tunable via ros2 param set: costmap_crossable_cost (20), costmap_inflation_radius (0 = off), costmap_inflation_cost (60), costmap_ground_tolerance (0.40 m).

Tests

The costmap geometry and the lane toggle are unit-tested — no camera, GPU or model weights required:

pixi run test
# or directly:
pytest src/zed_ml_inference/test/test_costmap.py -v

The tests cover grid dimensions and the origin pose, free/unknown/lethal classification, the ground-height filter (so a car or wall never becomes free space), out-of-range points being dropped rather than wrapping to the far side of the grid, both states of lanes_as_obstacles, the lane layer staying independent of the toggle, inflation, cost clamping, and the column layout shared with the lane/points/lanes PointCloud2.

The few tests that build an actual OccupancyGrid message skip themselves when nav_msgs is not importable, so the suite also runs on a plain dev machine with just numpy, opencv-python and pytest.