Anti-collision Robot Navigation Using Image Processing

[zaryn]

My idea is to build a simple robot, which is more likely a car,that can navigate itself without colliding with any barriers and without being controlled manually. I decide to use image processing as a method of barrier detection, which means my robot will be equipped with at least one pinhole camera to capture images while the robot is navigating. But, the critical problem is, I still couldn't find any method of barrier detection instead of edge detection method and also line detection method.. I don't think these method are good enough for detecting any obstacles or barriers.. Can anyone give some ideas?  ..For your information, I am doing my proposal for final year project, need to be submitted as soon as possible..

[hgordon]

One simple approach is to sample the colors of the floor surface in front of the robot and then use that to look for matches or mismatches as the robot moves around.  Pixels that match the floor color likely represent areas where it's safe for the robot to navigate, and non-matching pixels likely represent obstacles.

You can find examples of "floor matching" algorithms in Roborealm - see http://www.roborealm.com/help/Floor_Finder.php

[dunk]

need to be submitted as soon as possible..

if this is the case then i recommend you choose a different goal.
people put years of research into this goal with varying degrees of success.

if you are determined then i'd agree with hgordon that simple colour recognition is probably your best chance of getting a working solution quickly but you have to accept the limitations that that approach comes with. (mistaking floor coloured bits of wall for floor, etc...)

try searching this forum for the word "vision" as it's been discussed a fair bit all ready.
http://www.societyofrobots.com/programming_computer_vision_tutorial.shtml
jfgi.

good luck with it!
sorry to be less than optimistic about your chances.

dunk.

[Admin]

My idea is to build a simple robot . . . I decide to use image processing as a method of barrier detection

Vision is not simple

If your goal is to do object avoidance, why not use Sharp IR?
My Fuzzy robot scanned an area and made a map. Then it ran an algorithm on the map to decide the best clear path to take realtime.
At the very bottom of the page I have free code for you to download. That page also goes into detail about omni-wheel robot theory, of which you can ignore.

Otherwise, for vision, I also agree with hgordon.

[rgcustodio]

If you have the time you might want to research about "Stanley", Stanford's autonomous vehicle. It uses a several sensors (the 2000USD or more SICK laser range finders ones and a camera) to map the vehicle's current surrounding, which is used as inputs in the vehicles navigation system. But then again, this might be too complex.

As everyone has mentioned, image processing is complex and you will need to design for that. It will not be a simple task, specially if it'll be an embedded system. Hopefully, its just the proposal that needs to be submitted ASAP

[Admin]

Hmmm I feel dumb for not thinking of posting this before:
http://www.societyofrobots.com/robotforum/index.php?topic=1173.0
Brandon used sonar, and his design is probably the simplest you can make for your goals.

[zaryn]

I do noticed that image processing is too complex..But we are not allowed to use any sensor for this purpose of barrier detection, as the previous students have used IR for barrier detection for automated vacuum..so we need to find other method that might enhance the previous project..or at least,it would not be the same as that previous project.. As we are proposing something similar to it..Actually, we are doing this proposal for our final year project that will be implemented during our final semester soon.. so it's not just a proposal.. ..

[hgordon]

You can get surprisingly good results with the floor matching approach.  The main challenge is to develop an understanding of the range of colors the robot's video sensor will likely encounter, and appropriately set up your thresholds.  If the lighting is controlled, that greatly simplifies matters.  However, the video sensor will have problems dealing with reflective surfaces. 

Again, I'd suggest experimenting with the RoboRealm software just to see what works and what doesn't.  You can later translate that knowledge into your own software.

[Brandon121233]

I think in robotics it is also useful to recognize what sensors are good at and what they are not good at. Cameras are very good for color and object tracking, as well as path finding. And yes, you could probably use a camera to avoid simple objects, but if thats your only goal, why spend the time and money to do something that a simple IR rangefinder can do just as well and probably better. Remember this: Do not use a cannon to kill a mosquito! It will make your life easier.

[Admin]

Amen!

[Nyx]

Easy...

Use two identical cameras, regularly build a depth map from the disparity in the two images, and you can then extrapolate what is an obstacle based on the visual distance of the depth map samples and simple geometry.

[hgordon]

Do not use a cannon to kill a mosquito!

I'd argue the opposite - why avoid learning how to use the sensor that is the primary navigation means for the majority of known lifeforms ? 

A video sensor allows you to emulate a variety of other simple sensors for functions such as line following, light metering and basic color detection.  Combined in stereo or with a laser pointer, they can measure distance.  Then as you get more sophisticated, you can detect and recognize shapes, faces, characters, objects, etc.  You can start simple and build from there. 

Anyhow, that's my $0.02's

[Brandon121233]

read my reply carefully!

but if thats your only goal,

if in fact you wished to expand and use the camera for other stuff then sure do it, but using a $100 item and hours of coding SOLELY for the purpose of obstacle avoidance... thats a bit ridiculous for me.

Secondly, I must point out that even a simple fly has thousands of times more processing power and God given instinct to use its eyes when compared to a feeble man made robot... At this point in image processing for robots, I think its almost like comparing "apples to oranges". That is solely my opinion, so take it as you please.

[hgordon]

but using a $100 item and hours of coding SOLELY for the purpose of obstacle avoidance... thats a bit ridiculous for me.

I don't think we know enough about his project to assess the relevance of any particular approach.  All we know from his original statement is that he decided to use image processing, but hadn't selected an appropriate technique, and from a later statement that he couldn't use IR or a sensor for barrier detection, though in fact the robot has to use some kind of sensor to effectively interact with the world.  So a clearer definition of the task or the project constraints would be helpful in securing relevant advice.

[zaryn]

My point of using image processing as a method of obstacle avoidance is just to use another method instead of IR sensor..we are not allowed to use IR as our seniors had used that sensor for the purpose of obstacle avoidance..They had developed an automated vacuum with collision avoidance system,and they used IR sensor and limit switch to make the vacuum detect the barriers and turn the direction of navigation after the barrier is detected. for this project, they had faced a problem that the vacuum failed to turn the direction of navigation once it collided with the barrier. Our group decided to improve the obstacle avoidance system, but it must be using a different method..The advantage of our project compared to them is that our robot will navigate and detect barriers without need to collide to that barriers.. I never made a further study on IR sensor actually...and I didn't know what exactly made the vacuum failed to navigate smoothly..but our lecturers suggested to use image processing as to detect barriers..its really hard to pass this proposal..and this is our third time we proposed our project actually..I really hope this would be the last..

[zaryn]

Easy...

Use two identical cameras, regularly build a depth map from the disparity in the two images, and you can then extrapolate what is an obstacle based on the visual distance of the depth map samples and simple geometry.

can u explain more bout this?

[Admin]

http://www.societyofrobots.com/programming_computer_vision_tutorial_pt3.shtml#stereo_vision