I am testing several methods for finding region of interest in hand gesture. in opencv for example I found some methods like camshift (for tracking a interest object), some background extraction methods (MoG, MoG2, ..) which specially are used in video to subtract background from foreground, which can also be used when we have hand as an object in a video with a complex background. and also GrabCut and backproject methods which can be used for hands posture in a static state. Contours, edge detection or skin methods are some other approaches for detecting hand in an image or video. And lastly I found that haar cascade can be used as well. I want to know that for passing from this stage, which algorithm is the best choice, considering that I use images with complex background. some algorithms like Grabcut or backproject were good but the most important problem was that I should manually specify some regions as foreground or background and this is not what it should be. After choosing a method for roi, generally what are the most important features in hand gesture recognition? for extracting features which method is your suggestion? that can work well with one of the general classifiers like svm, knn, etc to classify an specified image.
Thank you all for taking your time
Skin detection on the base HSV is fast method. Of course, you must then check whether it is hand or not.
For hand gesture detection, (1) dividing your training data to a set of patches (2) Represent these patches with a discriminative feature set (3) Training a binary-classifier for detection which patch is contain the hand properties. (4)Applying the classifier to your test frame (or images) for finding those patches which are contain hand as ROI.
Now, you can model and analysis the ROI for recognizing the gesture.
As i know, the best way for this work is using the Convelutinal NeuralNetwork.
it help to share the features which is used for ROI detecting, also for gesture recognition. This way leads to very fast run-time.
We are working on human posture recognition. For detecting hand, we propose a modified haar like descriptor called Internal Haar and for recognition, it seems that improved kernel descriptor works well.
If you are interested, you can see our paper on FG 2015. The code is available for research.
Try haar like feature with Random forest. It would fast and accurate.
scikit learn has implementation of random forest and Haar like feature are not that bad.
You can even train same pipeline for the regression of hand position.
Skin color segmentation would be a good choice (based on HSI or YCbCr).
how to follow hand gestures with the WebCam camera using image processing
how to follow hand gestures with the WebCam camera using image processing
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence