Vehicle Safeguarding
Description

Laser range datas augmented with color

LADAR and high resolution cameras

In order to achieve the high degree of reliability required of the perception system, we have chosen our sensors so that they provide complementary information that can be exploited by our higher level reasoning systems. To correctly fuse information from the cameras, the laser range finders and the position estimation system we have developed precise multi-sensor calibration and time synchronization procedures.

We implemented feature extractors that analyze the images in real time and extract color, texture and infrared information that is combined with the range estimates from the laser in order to build accurate maps of the operating environment of the system.

Since our perception system had to be easily adaptable to new environments and operating conditions, hard coded rule-based systems were not applicable to the obstacle detection problems we were analyzing. As a result, we developed machine learning for classifying the area around the vehicle in several different classes of interest such as obstacle vs. non-obstacle or solid vs. compressible. Novel algorithms were developed for incorporating smoothness constraints in the process of estimating the height of the weight supporting surface in the presence of vegetation, and for efficiently training our learning algorithms from very large data sets.

The initial system, installed on the 6410 John Deere tractor, has been demonstrated in several field tests. We are currently focusing on a small stand-alone perception system that uses cheaper sensors and could potentially by used as an add-on module for several existing types of agricultural machinery.

NREC Software Version Controlled By BitKeeper.