250 likes | 396 Views
Applications in Real-time 3D Tracking. John A. Bender. Collaborators: Pietro Perona Luis Goncalves Ken Goldberg Karl Chen Ilan Lobel Steve Nowlin. Inspiration. Application. (image: Lorenzo Wang Graphics). Tracking Problem. Camera Design. Orange Micro iBOT IEEE 1394
E N D
Applications in Real-time 3D Tracking John A. Bender Collaborators: Pietro Perona Luis Goncalves Ken Goldberg Karl Chen Ilan Lobel Steve Nowlin
Application (image: Lorenzo Wang Graphics)
Camera Design Orange Micro iBOT IEEE 1394 (FireWire) webcam - $100 custom modifications
Camera Design Orange Micro iBOT IEEE 1394 (FireWire) webcam - $100 custom modifications • 3-port FireWire PCI card has total bandwidth of 400 Mb/s - $100 • 320x240 pixels, 3 color channels, 8 bits/pixel, 30 fps, 3 cameras • = 166 Mb/s • isochronous sampling?? • processing time??
Camera Modeling real-world image lens projected image
Camera Modeling real-world image refraction boundary lens projected image
focal length principal point Camera Modeling parameters of interest… camera matrix: Fx 0 Px 0 Fy Py 0 0 1 point on image plane: [ (x-Px)/Fx (y-Py)/Fy 1 ] (homogeneous coordinate)
3D Modeling transformations between cameras are modeled with rotation/translation components rotation vector: [ Rx Ry Rz ] rotation matrix: X0X1 X0Y1 X0Z1 Y0X1 Y0Y1 Y0Z1 Z0X1 Z0Y1 Z0Z1 translation vector: [ Tx Ty Tz ]
3D Modeling transform rays into a common coordinate system and intersect transformation for a homogeneous 3D point (z=1): X0 = R10*X1 + T10
Image Processing • Microsoft DirectX (DirectShow) interfaces with hardware • image callback executed on every frame with timestamp, • each camera runs in a separate thread • background modeling / foreground masking with • Intel Performance Primitives optimized image processing libraries • ellipse fitting using Intel Open Computer Vision libraries • triangulation including ellipse size • fish determination • TCP network streaming • executes at ~20 Hz on a Dell 2.8 GHz Pentium 4 with • 1.5 GB of 533 MHz DDR SDRAM
Path Determination each point is a 3D point which could contain a fish each column is a different time point
Path Determination draw paths from each point in first frame to each point in second frame
Path Determination draw paths from each point in first frame to each point in second frame
Path Determination extend each path to all possible positions
Path Determination extend each path to all possible positions cull paths with too-low traversal probability
Path Determination add dummy node with fixed transition probability in case of disappearances
Path Determination add dummy node with fixed transition probability in case of disappearances place dummy node at last known position
Path Determination for each ending node, choose the unique path that minimizes traversal probability
Rendering • virtual fish tank • each virtual fish follows the position of a real fish • Newtonian motion model for virtual fish ensures smooth movement • camera position in virtual tank set to head of gold fish
Results – from Vision… Results – from Vision to Reality