Cognex patents obstacle-detection system
In January 2004, Sanjay Nichani of Cognex (Natick, MA, USA; www. cognex.com) filed a patent describing a three-dimensional (3-D) machine-vision obstacle-detection system (US Pat. #6,678,394 B1). The system uses a Triclops stereo camera from Point Grey Research (Vancouver, BC, Canada; www.ptgrey.com) with a Cognex 8100 frame grabber to convert two-dimensional (2-D) video pixel data into 3-D point data to calculate the closest point from a mobile vehicle to the nearest 3-D objects.
According to Art O'Dea, patent attorney with Cognex, the patent evolved from a development Nichani was working on a few years ago to apply machine vision to control automated industrial vehicles. These driverless fork-truck type of vehicles drive around factories to move material. "Although that particular product is not currently under development, certain aspects of the technology are being used in other machine-vision applications using 3-D image data," he says. These include the development of security-based systems and possibly vision-based robotics control.
After acquiring 2-D images of a scene, Cognex's patented machine-vision system generates a set of 3-D data points that are compared with a reference model of the scene. The data are used to calculate to the shortest distance vector from the object in the scene and the target vehicle.
To implement obstacle recognition on a mobile vehicle, the system must first be trained to gather data about the vehicle itself. The data are then used to calculate distances from the vehicle to obstacles that will be encountered by the autonomous system. To accomplish this, stereoscopic images of the vehicle are first captured by a reference camera and processed to create a set of 3-D points.
"These 3-D points are obtained only at the boundaries of the objects and are classified as 3-D features," says Nichani. "These boundary points include the occlusion boundaries due to surface discontinuities and texture boundary points. Specific 3-D features are then derived using well-known edge segmentation processes and stereo algorithms," he says.
After the training phase, the same acquisition process is used by a stereo camera on the vehicle to generate another set of 3-D points about objects in the scene. These points are then compared with the set of points used during training, and for each 3-D point a result is generated that corresponds to the shortest distance vector from that point to the target vehicle. This vector is then used in a thresholding analysis to classify the 3-D points captured from the vehicle as target, obstacle, or background.
"In the present system," says Nichani, "we can generate real-time 3-D position information about objects in the viewed scene. And, if an optional 3-D segmentation algorithm is applied, this information can be resolved into 3-D objects. These 3-D objects are generated through a process of clustering of the 3-D data points into clouds that correspond to a 3-D object in the scene. Once a set of 3-D objects has been generated, filtering and comparison of the data is performed to detect obstacles in the vehicle's field of view. This object-recognition system identifies the position, shape, and size of objects in the vehicles path," he says.
The first Cognex product using this 2- and 3-D vision technology is the CPS-1000, a door-security system designed by Horton Automatics (Corpus Christi, TX, USA; www.hortondoors.com). The system detects and counts people as they pass through an access-controlled doorway.