DeSouza’s paper surveys the last twenty years of development in the field of vision for mobile robot navigation. One of the main issues that the author mentions is that of the definition of “progress” in the field of vision research. One example I can think of is Nilsson’s “A Mobile Automaton: An Application Of Artificial Intelligence Techniques” where the same image processing techniques developed in 1969 are being used today. It seems that the only change from this article was that the memory and disk space has gotten cheaper and faster. The important point is that how would one measure progress where an autonomous robot needs to seek out certain objects amongst clutter since different algorithms approach the problem differently giving the appearance that one works better than another, and yet in other situations that same algorithm would fail.
The main search in this field has been done with vision based navigation for outdoor robots, which is similar to Dellaert’s “Robust Car Tracking Using Kalman Filtering and Bayesian Templates” where it tracked cars in real time, Saripalli’s “Vision-based Autonomous Landing of an Unmanned Aerial Vehicle” where the helicopter could search and land on a helipad and Ettinger’s “Vision Guided Flight Stability and Control for Micro Air Vehicles” where the plane can figure out the horizon. The other field for vision based navigation is with indoor robots, like Nilsson’s paper. From what I have seen, indoor environments are more controlled and therefore the algorithms tend to produce results without worrying about filtering out environmental noise.
One interesting area with indoor vision based navigation deals with map building. Building an internal representation of the surroundings and using that information later to explore seems like a fascinating topic and issues surrounding it. The two main approaches with map building are occupancy-grid based (where a 2D grid of the area is created) and topology based (where a graph of the environment is created). I became curious about map building after I volunteered for an experiment on the Pioneer robot. They were testing out different interfaces for the robot and one question I remember asking was in regards to the mapping area in the GUI. They mentioned that with all the sensor data and other computations that the mapping piece was computationally expensive. The issue that I thought about at the time is that you need to consistently calibrate your sensors because the data you get back could potentially have errors. Over time these errors could create very wrong maps of the environment and if the robot is dependent on the proper map, this would significantly impede its ability to navigate its environment.
In the end, this survey paper represents the main issues with sensory intelligence, specific to vision based navigation. There seems to be an appreciation for prior knowledge and how this knowledge itself can be modeled from its environment. In the end, there seems to be an open field of study on solving these issues.
Reference:
- DeSouza, Guilherme N. “Vision for Mobile Robot Navigation: A Survey.” <http://vigir.missouri.edu/Publications/00982903.pdf>
