Berkeley Lab

Machine Vision Radiation Detection

The aim of the Machine Vision Radiation Detection System (MVRDS) project is to accurately and reliably detect the presence of a nearby, unknown radioactive source and to locate its position for the time that it is present within the range of detection. The main components of the system include NaI detectors to measure local radiation and cameras to determine the trajectories of moving objects in the area. Our task is to devise an efficient detection algorithm, effectively incorporating both types of data. Such a system could provide wide security benefits at ports, bridges, borders, and other areas where security is on high alert. The Bearing Group is currently working on both a 1-D and a 2-D detection system.

The 1-D Problem

The original 1-D problem was to detect the position of an unknown radioactive source on a train moving at constant velocity in a straight line of roughly 30″ — the region of interest. This setup consisted of a model electric train, two photoswitches, two NaI detectors, and two cameras. The photoswitches were placed one where the train enters and the other where the train leaves, and determined the train’s trajectory, since the velocity was assumed to be constant. Thus, the cameras were used only for imaging and did not provide much benefit to the detection algorithm itself. Though this setup was subject to a constant velocity constraint, and used photoswtiches which is excessive equipment given that we have cameras, it provided a foundational and working example from which we could make improvements in the future.

1-D Experiment Setup

Recent Work

Since the constant velocity constraint is an unreasonable assumption in many realistic situations, the next step is clearly to make a variable velocity system with only NaI detectors and cameras to provide the necessary trajectory information.

Recently, two undergraduates, Sabrina David and Ray Yamada, have been working with Dr. Barak Fishbain on motion detection. They have taken sample footage of the model train and applied a block-matching algorithm to detect motion in the video.