Introduction to Programming (IUT dpt Info, Polytech GE3)
Database (ITU-dpt STID)
OMGL (IUT dpt Info)
ASR3 (IUT dpt Info)
Tutored project first year of DUT
Recognition of sign language
PFE fifth year Polytech'Nice
The goal of the project is to develop a search engine in large image databases, for which the interaction between the user and the motor is controlled using a Microsoft Kinect. Inspired by what is done for the handling of new flat screens (http://www.youtube.com/watch?NR=1&feature=endscreen&v=5C1nADiC6OE) developments have allowed to reproduce these features with a Kinect (http :/ / www.youtube.com/watch?v=bCuetTWdaJQ).
In this project we will adopt the same type of man-machine interface for a user to search a database of images with simple gestures and find all pictures of cats or bus or other that are in the database .
To do this, it will implement machine learning algorithms to classify permettronnt image database (we will use it for the study of SVMs made progress but we can of course also consider the Randoms Forests). It will extract visual features (color, texture, edges ...) to be input SVMs or Random Forests.
Finally, it will implement different strategies for interactive otpimiser search based on user interactions with the engine. We hope to have a close result http://retin.ensea.fr/ but driven by a kinect.
Many extensions are possible, such as determining a region in an image with your fingers to find all the images that have similar regions, draw a shape found in the images ...
The goal of the project is to develop a system for the recognition of human activity. Whether kept at home for the elderly, assistance to people with disabilities, or for analysis and indexing of video data, this area of research and industrial development is booming. Evidenced by the summer school in Sophia Antipolis in early October "Human Activity and Vision Summer School" which brought together international researchers in the field (http://www.multitel.be/events/human- activity-and-vision-summer-school/home.php) or "testing" the Multimedia Grand Challenge for 4 years (http://sigmm.org/records/records1201/featured04.html).
If the last 4 years, some very interesting work has been done by experienced researchers (http://www.di.ens.fr/ ~ Laptev / actions /) with results often bluffing on media data, the problem is compounded when we attack the recognition of action "live".
By exploiting the information potential of the Kinect, it is definitely possible to improve the current results, it still must be able to learn from heterogeneous data provided by the Kinect (video, 3D skeleton ...) to extract a relevant global information. This is the goal of this project.
To do this, it will implement machine learning algorithms that will classify the acquired video (we rather rely on approaches Randoms Forests that will merge the intermediate decisions and combine the decisions of heterogeneous data). It will extract visual features (motion, body parts, 3D info ...) to be input Random Forests.
Many extensions are possible, such as creating a software dance classes that can recognize if the sequences have not been observed or performed ... Then you can give free rein to your imagination.