Results

The training set contains a total of 70 different samples, which is the result after adding five samples for each of the 14 signs from the default dictionary. These training samples are executed by the same user at the same position.
A set of test samples executed by four different users that differs in size is built to find the accuracy of the system. For every user, three different samples are collected which results in a total of 168 test samples.



After testing the system for its different configurations and approaches (i.e. which are the most meaningful futures to describe a joint, weight applied to each of the four used joints, and which is the used classifier), the results show that the best configuration of the system reaches an accuracy of 95.24%.

Conclusions

The initial goal has been satisfied since the use of the Microsoft Kinect XBOX 360TM with the combination of the developed descriptor and classifiers has been shown as an efficient way to solve the sign language translation problem.
As a future work, it will be interesting to track the finger positions so that the signs from the dictionary will belong to one of the existing official sign languages.