Lecture Slides

The Mapping Problem

• Last week we talked about localization, which is the movement around known environment.
• What if we don’t know the environment?
  • need to first build an internal representation from sensor data.
  • we’ll talk about tools for this and algorithm using these tools

Perceiving the environment

All of these give you single point and quantized information → point measurements of structure.

• Active range/depth sensors
  • limitations:
    • fails outdoors/direct sunglihgt
    • limited operating volume
    • power hungry
• Multi-view Triangulation (passive sensing)

Point Clouds

Each observation provides many "point" measurements of structure. These are stored in an unordered "point cloud" (sensor_msgs/PointCloud2). This format is similar to sensor_msgs/LaserScan but uses a different representation. Each point can hold a range of values (x, y, z, rgb, curvature, etc.). ROS internally uses the Point Cloud Library.

import sensor_msgs.point_cloud2 as pc2
for p in pc2.read_points(point_cloud, field_names= (“x”, “y”)):
	print(p[0], p[1])

Building a model

1. Each step, move by odeometry
2. measure new point cloud or convert laser scan
3. transform from base_link to odom frame
4. merge with existing point cloud model.

• Can run into memory constraints as we build up more point clouds
• Occlusion issues that distort the point cloud.
• Drift/noise can accumulate