I wish to attach two cameras to the inside of a car's windshield (preferably through the use of suction cups) looking out the front of the car. They will be arranged to return a stereo view at some distant point. This will be for detecting objects in front of the car. The car will be moving at normal speed (from 0 - 65kph). Are there any low cost cameras that could effectively be arranged in this fashion?
i guess the current standard cameras are the GoPro HD Hero Cameras. We are using them for many different applications and they come with different plugs and mountings. we even put them outside on the car with suction cups.
they are standalone, recording for up to 4 hours with full HD resolution
however they have a very wide field of vision, thus distortions which may be disruptive for your stereo vision task
does the camera data need to be computed in real-time or do you just want to collect training material?
GoPro HD cameras suffer jelly effect caused by rolling shutter, so I would recommend not using them in a moving platform.
Any global-shutter camera would do the job.
Keep in mind that you must:
-Calibrate the cameras so that you know the relative R and t and do the stereo rectification.
-Synchronize the cameras: so they capture images of the same lapse of time.
Perhaps your best option is to buy a pre-calibrated stereo rig, such as Bumblebee from PointGrey!
Check out the Kitti Car to get an idea of the amount of work required in setting up such a car.
http://www.cvlibs.net/datasets/kitti/setup.php
They also use PointGrey cameras, but not the stereo ones. This is probably because The BumbleBee Xb3 may not have a wide enough baseline longer range applications.
Thanks guys. The GoPro HD's certainly are reasonably priced but as mentioned by Marcos, may suffer from too much 'jelly effect'. The BumbleBee looks more professional but then again is much more expensive. I need to do more research. One thing I need to make sure of, I'm assuming I need to use stereo vision as I wish to detect objects (0 - 30m out) such as pedestrians and road artifacts.
Hi,
At 30m range i think you will struggle with 3d object recognition using stereo vision to obtain the reconstruction (unless you have loads of money). This is because disparity is inversely proportional to depth w.r.t the focal length and baseline of the fixed rig and the discretization error increases rapidly as depth increases. At 30m with a bumblebee XB3 the theoretically most accurate you can get is around 4m! So two points next to each other, say on the same object, 30m away, could be 4 meters apart in you estimation (when using disparity->depth stereo vision).
http://www.ptgrey.com/support/kb/data/modifiedstereoaccuracy.xls
Cheers
Paul
Thanks for all the replies. After doing some camare research, I'm now curious why a configuration of raspberry pi cameras:
http://www.amazon.com/Raspberry-Model-Revision-Camera-Module/dp/B00GNA0SZS/ref=sr_1_8?s=electronics&ie=UTF8&qid=1387812073&sr=1-8&keywords=raspberry+pi+camera+kit
following the techniques in this paper:
http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=4653062&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D4653062
would do the trick?
The pi camera is 5MP and captures 1080p/30 video. The paper describes a configuration of cameras setup to capture objects and track them. At the moment I don't see any downside to trying this. As of today, a pair of the cameras would run about $140. Thoughts?
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing