Vision Software for the RoboRoos

Exhibitor: Leighton Carr

Supervisor: Gordon Wyeth

Research Group: Complex and Intelligent Systems

Industry Sector: Sports / Leisure

The RoboRoos vision system is designed to provide reliable, accurate, real-time information to the intelligence system. Due to the fast pace of the RoboCup F180 league in which the RoboRoos compete, the work of this project included improvements to the reliability and speed of the system. The following paragraphs give details on the workings of the vision system. For more information about RoboCup or the F180 league, please visit: http://www.robocup.org

The RoboRoos’ vision system uses two Basler A301fc firewire cameras that work at up to 640 x 480 resolution at 80Hz. Both cameras are connected to a 1.6Ghz Pentium4A PC via a firewire hub. Two image capture threads are used, although images are processed serially. The cameras transmit a proprietry RGGB format, that is converted to standard RGB using lookup tables. The standard conversion process has been modified to merge four adjacent pixels such that the resolution is halved in both dimensions. This gives a whole field resolution of 480 x 320, discounting the overlap between the two cameras.

Colour thresholding is performed on every second pixel in both dimensions, resulting in a thresholded image of 240 x 160 (see figure). The field border is removed by region growing four white pixels in each corner and flagging each pixel grown as “field”. This process is made more efficient through ignoring four user defined rectangles on each of the corners, and a certain number of pixels along the sides of the field. It is necessary to perform this removal, as the RoboRoos marker system uses white markers. Any non-field pixels that border field pixels are thinned according to a neighbour threshold of seven. This is performed to remove the field lines.

Thresholding colours are calibrated by painting around predefined pixels in the UV colour space (see figure). Black and white are defined based on threshold values of Y in the same space. The thresholded image is then parsed for orange, blue and yellow pixels. Once found, the area around these pixels is re-thresholded at full resolution, and region segmented. Each coloured region is defined as an object and has its second moment of area calculated to determine stripiness. Objects are identified as robots provided they have sufficient pixels and are sufficiently circular. To prevent concave robot configurations from being re-thresholded twice, each coloured circle is painted after it has been examined.

The area around home colour objects is region segmented for white objects. The direction of the robot is found by averaging the direction of the most stripey white object, and the angle from the centre of the coloured circle to the centre of this object. The identity of each robot is determined by binary examination of the remaining white markers. The ball is located by examining orange pixels in an image that has not been thinned.

Robot locations are then passed to the perspective correction system. This is calibrated separately for each camera by clicking seventeen predefined points. The actual locations and the real locations are resolved through Tsai’s correction routine. The object locations and orientations are averaged across the two cameras. The RoboRoos’ vision system runs at 300 x 440 pixels at 70Hz. Robots are matched 99% of the time, with no false positives.

Thesis Document (PDF)