Image Analysis

Two independent algorithms were developed to automatically analyze the data and determine if a container is empty or not. One algorithm is rule-based and the other relies on statistical methods. The statistical algorithm can learn from operator feedback on uncertain cases. Because the algorithms tend to have different reasons for mis-identifying a container, we ultimately hope to create a voting scheme using both algorithms to reduce false alarm rates even further. Currently the algorithms are run independently for testing and characterization. For the remainder of this paper, only the rule-based algorithm will be discussed.
A critical first step in the automatic analysis of cargo scans is segmenting the image into various major portions, such as open space, trailer chassis, container floor/walls/roof, and cargo area. This step required a large amount of effort to teach the algorithms about slanted roofs (short trailers have an appreciable angle), walls that are not quite perpendicular, etc. Other real-world effects that had to be accounted for include container patches, tie-down hooks, and other routinely-encountered bits of hardware built-in to the container. Multiple trailer configurations are also common in some areas and the system control logic as well as the image analysis routines had to be programmed to handle them.