I am trying to build a software system which will sense the real word in 3D (and do some AI stuffs to detect obstacle etc). At present I am using the Microsoft kinetic / creative 3D depth camera for sensing the terrain in 3D. These sensors are Laser and IR based hence are limited in distance and the nature of material the rays are bouncing from (eg, sensing black surfaces are difficult than white one). I have also found hokuyu sensors, but they are very costly. Can anyone suggest some good and cost effective sensors to map the terrain in 3D?
********************************************************************************************
3D sensor Options:
1. Stereo vision:
Pros: Cost effective
Cons: large processing, complex algorithm
2. Hokuyu:
Pros: Reliable Data, Visual Studio supported.
Cons: Costly, less range
3. Velodyne (Google Car):
Pros: Range, Reliable, Outdoor Capability,
Cons: Cost (75,000 USD ) , Mechanical spinning Unit,
You can rapidly implement a stereo vision system, by using some high quality HD resolution webcams (around 80-100 euros for both). Once you mount them firmly on a beam metal structure (baseline distance can be regulated according to your required optical scene viewing angle), you need to perform a calibration method (http://www.vision.caltech.edu/bouguetj/calib_doc/) to extract the intrinsic and extrinsic parameters of your stereo system. These parameter you will use to rectify your acquired stereo images so a disparity extraction algorithm (i.e. SAD) can be later applied, in order to properly extract your 3-D depth image. The inter-related complexity of the SAD disparity extraction algorithm requires some adequate processing power, so it depends on you and your resources, whether you implement your algorithm on a PC, or an FGPA device.
The main problem of the stereo vision is that it is complex to get a high frame rate. If you do not have this requirement, this is a good choice for you.
I would say it really depends on the resolution you want to achieve. If you are just looking for a simple top view terrain model, actually just depth combined with position from GPS or similar ..., you could also think of ultrasonic range sensors. We also have different hokuyu sensors, but even these costly sensors have a range limit. What range and environment/application are you looking for?
Look at the Google self driving car. It uses many 3D methods, and none of the 3 cars in operation have hit a person, in the many years of operation. Looks like Google is going to approval for the car in other countries, other than Silicon Valley. Long distance uses laser beam mounted on a gimbal, with servos to sweep the forward subtended angle. And sonar for closer. And video. Point is, just one system will not suffice for human safety. If you are looking at non human safety application, then you might use just 2 methods. One method will not be enough. There are many web sites that document. Try www.RCCS.org as a starting place. The Long Beach Robot Club is quite old, and their web site points to other resource laden web sites.
Hi Peter Benjamin, Christos Georgoulas and Miquel Massot-Campos
I truly agree with your answer, But google car uses the velodyne 3D sensor, which is priced at 75,000 USD ( Import and local taxes extra ! ). very good range . clarity and reliable..... but very costly (for me to start-up, i need a cost effective solution ).
As per Christos Georgoulas the stereo camera system is very cost effective but needs a lot of CPU. Presently trying to setup the hardware and give it a try. Thanks Christos for pointing the directions.
Thanks a lot for all you answers.
I have consolidated all the option in the Question description space.
Request all to add more information on available sensors and frameworks.
Regards,
- Badges
- Science topic
- Similar topics
- Social Psychology
- Attitude