Re: Re: Hand detection, PCA & tracking

Hi,
Think that would be a good idea use the haar-features to detect the hand 
(a single pose, maybe open hand) and then track the hand with the camshift.
I saw this way working and the results were very good.
To classify the gesture of the hand a very good classifier is the SVM 
(Support Vector Machines), you can find the implementation of it in a 
library called LTI-LIB.
I hope i've helped you...

Gonçalo Monteiro


yuvaltwig wrote:

>Hi,
>
>i'm looking to be able detect human hand.
>what would be the best bet ? using haarcascades xml file or camshift ?
>is there a xml file allready made ?
>
>thx in advance :)
>
>Yuval
>
>
>--- In OpenCV@xxxxxxxxxxxxxxx, "rjdpauw" <rjdepauw@xxxx> wrote:
>  
>
>>Hello Juan & Jeff,
>>
>>I have currently built a hand detector that uses haartraining combined
>>with a particle filter for the tracking. The next step is to do the
>>gesture classification. Since I only want to differentiate between 2
>>gestures I'm thinking of PCA using two eigen object means.
>>
>>Greetings,
>>
>>Rob
>>
>>--- In OpenCV@xxxxxxxxxxxxxxx, "juanwachs" <juan@xxxx> wrote:
>>    
>>
>>>Hi Jeff!
>>>Well, I used once 'TemplateMatching" and also got terrible bad 
>>>results. But you can check yourself to know whether is your bug, or 
>>>it is only a bad idea to use the TemplateMattching. To do this, try 
>>>to match between the original image that you are looking for, and the 
>>>same image. You must get a value of '1', which is perfect matching. 
>>>Then, if use use it with test images, you will get values from 0 to 1 
>>>according to the similarity.
>>>Again, I got very bad results, when I used it in the past.
>>>May be a better idea is to extract some features from each sample, 
>>>and feed a K-means algorithm. I think that this will be a better idea!
>>>Take care,
>>>Juan
>>>
>>>--- In OpenCV@xxxxxxxxxxxxxxx, "Jeff Ng" <csisjeff@xxxx> wrote:
>>>      
>>>
>>>>Juan,
>>>>
>>>>I am doing similar things.
>>>>I try to use template matching(provide by opencv) instead of c-means
>>>>or k-means. Does it sound reasonable? What do you think?
>>>>
>>>>Also, have you tried cvMatchTemplate. I got a small area, say x, 
>>>>        
>>>>
>>>from
>>>      
>>>
>>>>a pic, say X, and try to match x with X using the function. But the
>>>>result floating point numbers all give the same value. I cannot get
>>>>any information from the result. Do you know where is my problem?
>>>>
>>>>Thanks.
>>>>
>>>>Jeff
>>>>
>>>>--- In OpenCV@xxxxxxxxxxxxxxx, "juanwachs" <juan@xxxx> wrote:
>>>>        
>>>>
>>>>>Hi Robert,
>>>>>Well, more or less I'm trying to do the same: track hand and 
>>>>>recognize gestures. There are two approaches that I think may 
>>>>>          
>>>>>
>>>achieve 
>>>      
>>>
>>>>>this goal:
>>>>>1) Track the hand by skin color or movement or both together (I 
>>>>>          
>>>>>
>>>use 
>>>      
>>>
>>>>>camshift which works nice)
>>>>>2) Once the hand is resting, call to the recognition module. For 
>>>>>this, you need to: segment the whole hand from the background, 
>>>>>          
>>>>>
>>>and 
>>>      
>>>
>>>>>then call to a classifier. I'd suggest k-means or fuzzy c-means.
>>>>>
>>>>>The second approach is to use Haar classifier. It will find the 
>>>>>gestures, withuot motion or color information.  
>>>>>
>>>>>I already did the approach 1, and find that it is relatively 
>>>>>          
>>>>>
>>>fast, 
>>>      
>>>
>>>>>but inaccurate.  
>>>>>The seond approach should be more accurate, but probably it will 
>>>>>          
>>>>>
>>>be 
>>>      
>>>
>>>>>slower (since you have to train N classfiers, as the number of 
>>>>>gestures that you have).
>>>>>Probably the best is to combine the two approaches. 
>>>>>I'll try myself the method of Haar classifier, but this may take 
>>>>>          
>>>>>
>>>some 
>>>      
>>>
>>>>>time.
>>>>>Good luck!
>>>>>Juan 
>>>>>
>>>>>
>>>>>--- In OpenCV@xxxxxxxxxxxxxxx, Robert Jan de Pauw <rjdepauw@xxxx> 
>>>>>wrote:
>>>>>          
>>>>>
>>>>>>Hi all,
>>>>>>
>>>>>>I am currently using the OpenCV library to build an advanced 
>>>>>>            
>>>>>>
>>>user
>>>      
>>>
>>>>>>interface using hand gesture recognition. I already have a 
>>>>>>            
>>>>>>
>>>detector 
>>>      
>>>
>>>>>up
>>>>>          
>>>>>
>>>>>>& running thanks to the haartraining 'suite'. It is currently 
>>>>>>            
>>>>>>
>>>able 
>>>      
>>>
>>>>>to
>>>>>          
>>>>>
>>>>>>detect only one position of the hand. At first I thought of 
>>>>>>            
>>>>>>
>>>>>building a
>>>>>          
>>>>>
>>>>>>detector able to detect multiple hand gestures but that didn't 
>>>>>>            
>>>>>>
>>>>>really
>>>>>          
>>>>>
>>>>>>seem to work out.
>>>>>>
>>>>>>My strategy is the following:
>>>>>>
>>>>>>1. Detect the hand
>>>>>>2. Track the hand to make up for undetected hand gestures
>>>>>>3. Distinguish between gestures through euler-distance & PCA (or
>>>>>>perhaps kmeans ?)
>>>>>>
>>>>>>While the OpenCV library seems to support tracking, PCA and 
>>>>>>            
>>>>>>
>>>kmeans,
>>>      
>>>
>>>>>>information on these subjects is lacking. I therefore have a 
>>>>>>            
>>>>>>
>>>couple 
>>>      
>>>
>>>>>of
>>>>>          
>>>>>
>>>>>>questions:
>>>>>>
>>>>>>1. Suggestions for improving the hand detector (i.e. enabling 
>>>>>>            
>>>>>>
>>>it to
>>>      
>>>
>>>>>>detect multiple gestures).
>>>>>>2. What's the best tracker to use and how to use it (e.g. 
>>>>>>            
>>>>>>
>>>sample)?
>>>      
>>>
>>>>>>(Condensation is not supported anymore ?)
>>>>>>3. How to use PCA (e.g. sample)?
>>>>>>
>>>>>>I know that I'm asking a lot and I don't expect answers to all 
>>>>>>            
>>>>>>
>>>of my
>>>      
>>>
>>>>>>questions. Any help would be appriciated.
>>>>>>
>>>>>>Thanks a lot in advance,
>>>>>>
>>>>>>Rob
>>>>>>            
>>>>>>
>
>
>
>
>
>
>Change settings: http://www.yahoogroups.com/mygroups, select
>   Get Emails (get all posts)
>   Daily Digest (one summary email per day)
>   Read on the web (read posts on the web only)Or Unsubscribe by mailing 
> OpenCV-unsubscribe@xxxxxxxxxxxxxxx
> 
>Yahoo! Groups Links
>
>
>
> 
>
>
>
>
>  
>



Change settings: http://www.yahoogroups.com/mygroups, select
   Get Emails (get all posts)
   Daily Digest (one summary email per day)
   Read on the web (read posts on the web only)Or Unsubscribe by mailing 
OpenCV-unsubscribe@xxxxxxxxxxxxxxx