Year
2024
Abstract
Detecting and localizing nuclear material containers (NMCs) is necessary before artificial intelligence and robotics techniques can be efficiently deployed to reduce the burden of routine and laborious assaying and inspection tasks. The presented work explores a solution to this problem with the goal of mapping and counting NMCs. A handheld sensor system recorded various arrangements of AT-400 barrels at facilities in the Nevada Nuclear Security Sites. Imagery from a camera was processed with object detecting and instant segmenting neutral networks. Lidar and accelerometer data enable the creation of a three dimensional model of the inspected facility and to simultaneously calculate the movement of the system through the model. The two data modalities are fused to create a holistic understanding of the placement of individual containers. To assess the proposed algorithm’s performance, the predicted number of NMCs is compared to the actual number of NMCs present in a scene. The predicted locations are compared to a human labeled ground truth. Including scenarios that were designed to be particularly challenging, it is found that the algorithm is capable of identifying the number of containers correctly with a mean absolute percentage error of less than 10%.