Which ways to acquire range data do you prefer?

Personally, I use a tape measure. However, I''m not a computer agent. So, are you asking people what algorithms they recommend to extract range data from 2-D image data. Or, are you asking people what hardware sensor they would use to obtain range data for a robot implementation... or perhaps what software simulation of a sensor they would use to generate range data?

In other words, please be more specific about the information that you require. Perhaps by giving as a little more background on the problem you are investigating, we will be better able to describe to you some useful methods or algorithms to solve your problem.

Cheers,

Timkin

I want to know how to extract range data from 2d-image if it''s possible, or to advice me which range sensor is better to buy...

I need this for object recognition problem.

Generally it''s not possible to get range data from a single 2D image. You would need at least two images, and then you could measure parallax to determine ranges. Usually this is not a trivial task though. It all depends on the environment in which your system will operate.

You could always use sonar or a laser range finder instead of relying on the camera to gather range information. This would be much more accurate and require a lot less processing time.

If you do decide to go with a multiple image based solution you''re going to have to start working on some classification algorithms. You''ll need to be able to associate features in one image with their corresponding feature in the other image. Do a search on "Mean Shift" and "Image Segmentation".

Best of luck,
Will

If you want object recognition are you sure you want a range map? I''m not sure that you''ll get a sufficiently detailed depth map to be able to determine the features of the object to an degree of detail. Plus even today calculating depth maps isn''t a fast process.

Plus consider could yourself do it - if I gave you a greyscale depth map of a complex scene could you identify the objects in it? Then add blur, and noise and reduce to a computable resolution like 320x240.

On the other hand if you want to recognitise certain objects from a picture that is a completely different problem - and liable to be easier since a lot of object have lots of texture detail that the range map effectively throws away.


Keywords:

Camera Calibration
Stereographic vision
Monocular vision

Put them in google or citeseer

of course you could always use the depth map and keep the approximate pixel color for each unit in the map, hence giving you both the depth map and the texture of the scene.

just a thought.

- jeremiah

inlovewithGod.com

Of course, you could go for a data fusion solution: combine a 2D image analysis with data from a range finding sensor, such as a laser range finder. The sensor gives you an approximate depth map of objects within line of sight, which can be used as a basis for classification. I.e., utilise the range data to give an intial depth classification to pixel regions in the image and then refine the segmentation into more accurate segments, possibly with a split and merge algorithm and utilising another information measure (like minimum variance, minimum message/description length, etc). You could then utilise this segmentation to build an approximate geometric model of the environment (by turning the 2D image with depth segmentation into an oct-tree decomposition of the surrounding space).

Cheers,

Timkin

[edited by - Timkin on September 27, 2003 4:40:20 AM]

quote:
...use the depth map and keep the approximate pixel color...

Well, yes, there are many ways to do it - and ideally you''d use static imagary, motion based clues (both of object and the camera) and depth image to produce a complete object.

To achieve this recognition with any sort of accuracy I think that you''ll need to use a hybrid technique.

However I would have phrased this ''use pixel colour and keep an approximate depth map'' as pixel colour is free and ''accurate'' whereas depth is computationally expensive and far more liable to error.

Any chance of getting some more info from the original post? Will this be a 2d static image or a camera feed? do you know what sort of image you are looking at? Does it have to run in realtime?

As some ideas though - without even going into real depths map calculations you can make educated guesses about the layers in the image based on region finding.

quote: Original post by Anonymous Poster
As some ideas though - without even going into real depths map calculations you can make educated guesses about the layers in the image based on region finding.

Also known as segmentation! The difficulty of course in this sort of problem when performed in the real world is that inference of the depth of an object, or the inclination of a plane within the image is correlated with the light source(s) present in the image. Thus, one must infer both the position of the light source(s) and the resulting light intensity map and use this to infer the object model within the field of view.

Of course, if you have sequential images from a moving camera, or a moving target, then the task is easier, since you can utilise temporal correlations to improve the inference.

Cheers,

Timkin

I think it all depends on the environment he wants this stuff to work in.

If you''re looking for buildings you could use a Hough transform to locate edges and vertecies. You should then be able to match the vertecies between various image sets, and interpolate the depth from vertex to vertex. This might actually work if you have 5 to 10 quality images of the same scene.

Seach the all-knowning Google for "photogrammetry". It could be what you''re looking for.

I didn''t know so many people in the AI forum were interested in image processing.

Cheers,
Will