- Ultrasound navigation- Basically robot will have mic mounted and it will be listening for a signal from the speakers. There will be four ultrasound speakers in each corner emitting sound. At a start robot will send by RF signal to first speaker to send sound and then will measure time and calculate distance. Then will do this process with second speaker and so on. When it will come to the end, it will take all distance and calculate where exactly is it on field.
But there are some problems. Animals can pick ultrasound and can be disruptive for them. Second, any obstructions such as walls in between or trees can produce fault calculations. Accuracy will depend on the quality of hardware as well as weather conditions which might be difficult to compensate easily.[/pre][/pre][/pre][/pre]
Ultrasonic barely works indoors
Outdoors I suspect there might be attenuation and odd multi-path reflections.
-RF signals: Same as ultrasound, but now with antennas. Electromagnetic wave are to fast to measure time, but you can detect signal strength. Use multi antennas around garden, which produce unique signals, so robot can pick each signal strength and calculate position base on a (1/r^2) distance power lost.
iRobot apparently is doing some research in this area:
http://www.wired.com/2015/04/irobot-lawnbot/I'm a noob when it comes to RF voodoo, but their idea seems very strange. Why 6.5 GHz in particular? Why 500 MHz wide span? You could have shitload of channels and huge bandwith in such wide band I suppose. Seems like an overkill just for triangulation.
-Camera:I believe this technique will work, but it is to hard to create good program that will be accurate enough. You can but camera in fix position somwhere in field and track robot this whey. (using some easy to track object on robot like red circle), but any obstacles, threes, bushes, walls,... will ruin it.
Another option is to use a camera on a robot and some well-known shapes/colors, and run simple image recognition around the garden. Use triangulation (perhaps rotating the web cam with a stepper motor) to figure out where you're at. But again to hard to pull off.
Your first scenario is pretty much what Vicon (
http://www.vicon.com/) system does. High speed high resolution IR sensitive cameras, IR reflectors and IR reflective markers on your robot. Very expensive system, probably because cameras use CMV2000 sensors. Also, I'm not sure if it would work outdoors.
I'm somewhat very familiar with the second scenario, doing a PhD in this area
You don't necessarily need particular shapes/colors - just generic image features (google: Harris corners, FAST detector, SIFT descriptor and dozens of others) that can be matched across images.
From there there are multiple ways of doing SLAM (
https://en.wikipedia.org/wiki/Simultaneous_localization_and_mapping). Monocular = one camera, where you match features from frame to frame while moving and from that you can localize yourself and estimate the map. Downsides: you have to move and scene (environment) should be static preferably. Examples: classic MonoSLAM and PTAM algorithms, Dyson 360 Eye robot (crazy camera there), Google Tango.
Stereo vision is more expensive, but more robust and practical. Automotive computer vision often uses stereo for these reasons. Even when the robot is motionless you can estimate feature locations based on two synchronized camera images. For example:
http://uk.mathworks.com/help/vision/examples/sparse-3-d-reconstruction-from-two-views.htmlHere I mentioned sparse (feature-keypoint) based systems, as these are more established and well known outside academia. Also faster and easier to run on an embedded platform. There are also more recent dense methods as well, estimating things of interest pretty much per pixel, but these often require a massive GPU
Side note: MS Kinect is awesome indoors, but rubbish outdoors. Its powerful laser projector is no match for the sun
Side note 2: Typical cheap cameras are rolling shutter and therefore it might be difficult to get algorithms described above to run properly.
-GPS:I heard, they sell some really accurate GPS, but price is to high. I believe that army must have some pretty precise system to navigate. Don't know much about it.
This might be the best option. RTK GPS (
https://en.wikipedia.org/wiki/Real_Time_Kinematic) can provide cm level accuracy and you would need two GPS receivers and a radio link. There is an open source library called RTKLIB and it supports popular uBlox and Trimble modules. Unfortunately, I don't have any hands on experience with such GPS setups.