Vision for mobile robot navigation Jannes Eindhoven
Contents Introduction [2] Indoor navigation Map based approaches [5] Map building [1] Mapless navigation [2] Outdoor navigation In structured environments [3] In unstructured environments [1] Summary [1]
Introduction Guilherme DeSouza Avinash Kak
Introduction[2] Summary of the developments of the last 2 decades. February 2002, thus not including latest developments Not all-comprising Gives examples of achievements
Indoor navigation – map based Acquire sensory information Detect landmarks Establish matches between observation and expectation Calculate position
Map based – absolute localization Initial position is unknown Multi belief system Known landmarks from a map Calculate the position, incorporating the uncertainty in the landmark locations Metric map
Map based – incremental localization Start position is known Uncertainty in position is projected in camera image Only use features in their expected image parts The position gets updated
Map based – incremental localization [2]
Map based – Landmark tracking Artificial landmarks Natural landmarks Geometric and even topological representations Example: NEURO- NAV
Map building Slow process Additional problem to localization Generating occupancy grid or topological map with metric representation at nodes
Mapless navigation No explicit map Storing instructions as direct association with perception
Mapless navigation – optical flow Corridor following Viewing sideways, measuring surface speed and proximity of both walls Direction determined by PID controller Problems with walls with little visible features
Mapless navigation - Appearance-based matching Memorizing the environment Associate commands or controls with these images Like a train with a movie as “track” Can be simplified by matching only vertical edges
Outdoor navigation Changing lightning is challenging Main application is car automation
Outdoor navigation – Structured environments Navlab's ALVINN Neural network with picture or Hough transformed picture as input Lighting and shadows are a problem
Outdoor navigation – Structured environments [2] Virtual camera images, extracted from the original camera image Red and blue contrasts Speed is required for automotive applications Hue / intensity images
Outdoor navigation - Unstructured Measuring local environment metrical Example: Pathfinder rover and lander
Conclusions In controlled environments a lot can be achieved with current knowledge In free or unpredictable environments, there is still a long way to go