-

hey viewer, we're moving!

We are currently transitioning to a new web system, so we are not updating this wikisite anymore.

The public part of the new web system is available at http://www.ira.disco.unimib.it


Vision-based Autonomous Driving

From Irawiki

Jump to: navigation, search

People participating to this project

The aim of this project is to develop a robust system to determine the position of a moving observer using different kind of sensorial inputs (IMU, Computer Vision, wheel odometry) as well as to detect and track objects in the world.

Contents

The stereo-head

First of all, to get the USAD, Urban Shuttles Autonomously Driven able to see, we need to create a stereo camera vision system. We will use two cameras PhotonFocus MV1-D1312-40-GB-12 with the following characteristics:

  • Resolution: 1312x1082 pixels
  • Interface: Gigabit Ethernet
  • Lens Mount: C
  • Image: Black/White
  • Shutter: Global
  • Pixel size: 0.008mm
  • Pixel aspect ration: 1

We will also use the following Pentax B618(KA) C20616KA lense:

  • Focal: 6.5mm
  • F-Number: F/1.8

With both cameras and lenses we have the following data:

  • Horizontal view angle (half): 38.92°
  • Vertical view angle (half): 33.66°

What we need

The aim of the installation of the stereo head is beeing able to "see" as near as possible to the fron of the cart and, also, more lateraly possible. We also have the need of beeing able to see people that are close to the cart and as well as people that are far from the cart. To do this we need a discretely wide FOV (Field Of View) of the cameras and horizontal inclination angle (on the X axis) of cameras itself able to match all the requisites.

As written before, cameras have a vertical FOV angle of 33.66°, so the inclination angle to view at infinite must be less than 33.66° The stereo head will be mounted at a height of about 1,75 meters from the ground and approximatively at 50cm from the vertical plane of the most forward part of the cart.

From the simulation made with Matlab (the m-file can be downloaded here and personalized to view results with different parameters), the best inclination of the cameras (referred to the horizontal plane and considering positive counterclockwise) is 25 degrees. This inclination make possibile to still view quite near the cart front and, at the same time, it's possible to view a full person 180cm tall at about 1 meter from the focal plane.

The minimum distance from the focal plane (referred to the ground Z=0) viewable from cameras is at 1 meter from the focal plane. The cameras will be mounted approximatively 50cm behind the most forward part of the cart, so we will be able to see an object on the ground at 50cm from the cart. We also must consider that in the front part of the cart will be mounted the laser-scanner equipment, so the distance of 50 cm is more than reasonable. Also, we'll be able to see a full person at only 1 mt from the cart and that's very usefull. Here some theoretical images from Matlab about the field of view with the cameras mounted at a height of 175 cm and with 25 degrees of inclination.

File:USAD cameras FOV perspective view.gif File:USAD cameras FOV lateral view.gif File:USAD cameras FOV top view.gif

Here some images captured with the camera placed with a tripod more or less in the same position where the stereo head will be mounted.

File:USAD cam test 1meter front.jpg File:USAD cam test 1meter fron 1meter lateral.jpg File:USAD cam test 10meter fron 5meter lateral.jpg

Creating the stereo-head

The mounting of the stereo head must be strong and stable enough to be sure that the cameras will not move during the normal use of the cart.

The stereo head support is made with an empty aluminium box 4mm thick. This make sure that the support doesn't twist under stress. Cameras will be screwed at this support with two M4 screws. To be sure that the position of the camera will remains unchanged, an ad hoc 2mm thick support has been made. This additional support is fixed at the cameras with six M4 screws each and it is anchored at the empty alluminium box with other six M4 screws each. In this way we can be sure that the position of the camera will not change.

The support on the Cart will be again the empty alluminium box 4mm thick anchored with three M8 screws at the top-front "roll-bar" of the Cart.

The connection with the two supports has been performed with a swivel joint bought from "ITEM". Using 3 swivel joints it is possible to define the pitch of the camera while, at the same time, keeping them on the same position.

{add photos}

Mounting the stereo-head

Image acquisition

It is important to have good images while performing disparity map computation. This mean that images must not be blurred (exposure time too high), must not be overexposed or underexposed. Furthermore, image from Left camera and from Right camera must be taken at the same time.

Image quality

Photonfocus cameras support the LinLog function. (...)


Power supply and consumption

The cameras need a power supply of 12V (+/- 10%).

The datasheet report a maximum power consumption of 5W for each camera, so for the stereo head we will need a maximum current of 830mA@12V (10W). A single 12V/7Ah battery will be able to run the stereo head for about 8 hours. It is also necessary to consider the ethernet switch power consumption at which the cameras will be connected {still no data for eth switch}.

Images sets

Hereafter some acquisitions sets from the stereo head.

EIV exibition (indoor)

These images has been taken during the EIV exibition @ Rho-Pero. The illumination is variable and move from artificial light to direct sunlight. Images are already rectified and each image has a post-processing via a personalized function. LinLog function has a personalized response curve to fit the best in each condition. Left and Right camera are software triggered. Camera settings are:

Exposure time: 10ms

Aperture: F1.8

Acquisition group 1

Snowy weather (outdoor)

These images has been taken on friday 17 december 2010 from about 16:00 to 17:30 in snowy conditions. The illumination source is firstly natural light, then dark with public illumination light. Images are already rectified and each image has a post-processing via a personalized function. LinLog function has a personalized response curve to fit the best in each condition. Left and Right camera are software triggered. Camera settings are:

Exposure time: 10ms

Aperture: F1.8

Acquisition group 1 - Acquisition group 2 - Acquisition group 3 - Acquisition group 4 Acquisition group 5 - Acquisition group 6

Sunny weather (outdoor)

These images has been taken on friday 13 january 2011 from about 13:00 to 14:00 in sunny conditions. Images are already rectified and each image has a post-processing via a personalized function. LinLog function has a personalized response curve to fit the best in each condition. Left and Right camera are software triggered. Camera settings are:

Exposure time: 10ms

Aperture: F1.8

Acquisition group 1

External Links

PhotonFocus software repository

Photonfocus software for linux (x86) for Redhat distro. Photonfocus tech support says it has also been tested under suse,debian,fedora and ubuntu but they don't give support for these distros

Photonfocus software for linux (x64) for Redhat distro. Photonfocus tech support says it has also been tested under suse,debian,fedora and ubuntu but they don't give support for these distros

Personal tools