El documento proporciona una descripción detallada de cómo funciona Kinect para el reconocimiento de movimiento. Explica que usa una cámara RGB, un sensor de profundidad y micrófonos para capturar datos sobre la posición, movimiento y voz del usuario. Luego, mediante un proceso de entrenamiento con aprendizaje automático basado en millones de imágenes, Kinect aprende a identificar y seguir las partes del cuerpo humano en tiempo real para generar un esqueleto digital del usuario.
7. Dispositivo que combina una cámara RGB, un sensor de profundidad y un array de micrófonos Cámara RBG para el reconocimiento de los tres colores básicos Sensor de Profundidad que permite “ver una habitación en 3D” El array de micrófonos detecta las voces y las aisla del ruido ambiental Caja negra de software que une todo y hace toda la magia ¿Qué es Kinect?
37. J. Shotton, J. Winn, C. Rother, A. Criminisi, TextonBoost: Joint Appearance, Shape and Context Modeling for Multi-Class Object Recognition and Segmentation. European Conference on Computer Vision, 2006 MSResearch: Reconocimiento de Objetos
38. Amplio campo de acción Pero poca “agilidad” Y no es real-time MS Research: Human Body Tracking R Navaratnam, A Fitzgibbon, R Cipolla The Joint Manifold Model for Semi-supervised Multi-valued RegressionIEEE Intl Conf on Computer Vision, 2007
39. Necesitamos un body tracker con All body motions… Allagilities… 10x Real-time… Formultipleplayers… … and it has to be 3D XBOX llama a MSR: Septiembre 2008
40. Paso 1: Recolección de información El equipo visita diferentes ubicaciones y se dedica a filmar usuarios reales de Xbox Hollywood motion capture studiogeneratesbillions of CG images MSR & xBox: Machine Learning
42. Identificar cada pixel asociado a una de las 32 partes del cuerpo humano Crear un cluster con las posibles configuraciones de “partes” que coincidan con las articulaciones Presentar la probabilidad más acercada a la realidad al usuario t=1 t=2 t=3 Indenticando el cuerpo
43. Millones de imágenescomoreferncias-> millones de parámetros de clasificación Very far from “embarrassingly parallel” Nuevo algoritomopara resolver árboles de decisióndistribuidos Utilizaciónmasiva de DryadLINQ Disponibleparadescargar Training Distributed Data-Parallel Computing Using a High-Level Programming Language M Isard, Y Yu International Conference on Management of Data (SIGMOD), July 2009
46. Architectura extensible Expert 1 fuses the hypotheses Arbiter Expert 2 Expert 3 probabilistic Final estimate Raw data Skeleton estimates Sensor Stateless Statefull 37
47. Sensor Mapa de profundidad Separación por jugador basado en el fondo Paso a paso para el reconocimiento 38 Clasificación de partes del cuerpo Identificación de “joints” Creación de “Skeleton”
56. Seguimiento de cabeza y manos 2 “Seguidores” (trackers) Seguimiento de cuerpo not exposed through SDK 43
57. El problema del seguimiento de cuerpo Classifier Input Depth map Output Body parts Runs on GPU @ 320x240 44
58. Entrenando a Kinect Comienza desde datos ground-truth Alineados con partes del cuerpo Es necesario entrenar a Kinect para trabajar con Poses Posición por escena Tamaño y formas del cuerpo 45
72. Compute P(ci|wi) pixels i = (x, y) body part ci image window wi Learn classifier P(ci|wi) from training data randomized decision forests Clasificando pixel a pixel example image windows window moves with classifier 50
74. Analizalasposiciones 3D del todaslaspartesidentificadas del cuerpo Genera unacolección (posicion, confidence)/parte Genera múltiplesopcionesparacada parte del cuerpo El trabajo lo realiza la GPU Paso 1: «Body» a «Joint Positions» 52
75. Basado en 3 modelos de “Skeleton“ El proceso se realiza en: Cálculo de distancia entre puntos conectados(relativos al «tamaño del cuerpo») Cercanía de los huesos con las partes del cuerpo Aplica además patrones para el «smoothness» Paso 2: «Joint Positions» a «Skeleton» 53
kinetic," which means to be in motion, and "connect," which means it "connects you to the friends and entertainment you loveNatural User InterfaceMaking Beginners Feel Like Experts
Play video if you have time and if people have not seen Kinect in action
Color VGA video camera - This video camera aids in facial recognition and other detection features by detecting three color components: red, green and blue. Microsoft calls this an "RGB camera" referring to the color components it detects.Depth sensor - An infrared projector and a monochrome CMOS (complimentary metal-oxide semiconductor) sensor work together to "see" the room in 3-D regardless of the lighting conditions. Complementary metal–oxide–semiconductor (CMOS) (pronounced /ˈsiːmɒs/) is a technology for constructing integrated circuits. CMOS technology is used in microprocessors, microcontrollers, static RAM, and other digital logic circuits. CMOS technology is also used for several analog circuits such as image sensors, data converters, and highly integrated transceivers for many types of communicationMulti-array microphone - This is an array of four microphones that can isolate the voices of the players from the noise in the room. This allows the player to be a few feet away from the microphone and still use voice controls.What comes in the boxKinect sensor for Xbox 360Power supply cableUser's manualWi-Fi extension cableKinect Adventures gameColor VGA Motion Camera 640 x 480 pixel resolution at 30FPSDepth Camera 640 x 480 pixel resolution at 30FPSArray of 4 microphones supporting single speaker voice recognitionKinect's software layer is the essential component to add meaning to what the hardware detects. When you first start up Kinect, it reads the layout of your room and configures the play space you'll be moving in. Then, Kinect detects and tracks 32 points on each player's body, mapping them to a digital reproduction of that player's body shape and skeletal structure, including facial details.http://electronics.howstuffworks.com/microsoft-kinect3.htmhttp://www.popsci.com/gadgets/article/2010-01/exclusive-inside-microsofts-project-natalKinect Software Learns from "Experience"Kinect's software layer is the essential component to add meaning to what the hardware detects. When you first start up Kinect, it reads the layout of your room and configures the play space you'll be moving in. Then, Kinect detects and tracks 48 points on each player's body, mapping them to a digital reproduction of that player's body shape and skeletal structure, including facial details [source: Rule].In an interview with Scientific American, Alex Kipman, Microsoft's Director of Incubation for Xbox 360, explains Project Natal's approach to developing the Kinect software. Kipman explains, "Every single motion of the body is an input," which creates seemingly endless combinations of actions [source: Kuchinskas]. Knowing this, developers decided not to program that seemingly endless combination into pre-established actions and reactions in the software. Instead, it would "teach" the system how to react based on how humans learn: by classifying the gestures of people in the real world.To start the teaching process, Kinect developers gathered massive amounts of data from motion-capture in real-life scenarios. Then, they processed that data using a machine-learning algorithm by Jamie Shotton, a researcher at Microsoft Research Cambridge in England. Ultimately, the developers were able to map the data to models representing people of different ages, body types, genders and clothing. With select data, developers were able to teach the system to classify the skeletal movements of each model, emphasizing the joints and distances between those joints. An article in Popular Science describes the four steps Kinect's "brain" goes through 30 times per second to read and respond to your movements [source: Duffy].The Kinect software goes a step further than just detecting and reacting to what it can "see." Kinect can also distinguish players and their movements even if they're partially hidden. Kinect extrapolates what the rest of your body is doing as long as it can detect some parts of it. This allows players to jump in front of each other during a game or to stand behind pieces of furniture in the room.
Depth sensor. An infrared projector combined with a monochrome CMOS sensor allows Kinect to see the room in 3-D (as opposed to inferring the room from a 2-D image) under any lighting conditions.
a 320×240 depth stream. Depth is recovered by projecting invisible infrared (IR) dots into a room. The way the optical system works, on a hardware level, is fairly basic. A class 1 laser is projected into the room. The sensor is able to detect what's going on based on what's reflected back at it. Together, the projector and sensor create a depth map. The regular old video camera is held at a specific distance away from the 3D part of the optical system in a precise alignment, so that Kinect can blend together the depth map and RGB picture for dynamic, on-the-fly green screening.
RGB camera. Kinect has a video camera that delivers the three basic color components. As part of the Kinect sensor, the RGB camera helps enable facial recognition and more.
Four different microphones allow Kinect to figure out where the sound is coming from
Multiarray microphone. Kinect has a microphone that is able to locate voices by sound and extract ambient noise. The multiarray microphone enables headset-free Xbox LIVE party chat and more.
Microsoft software. A proprietary software layer makes the magic of Kinect possible. This layer differentiates Kinect from any other technology on the market through its ability to enable human body recognition and extract other visual noise.
Micron scale tolerances on large componentsManufacturing process to yield ~1 device / 1.5 seconds
http://research.microsoft.com/en-us/projects/DryadLINQ/DryadLINQ is a simple, powerful, and elegant programming environment for writing large-scale data parallel applications running on large PC clusters.