Sparse detector sensor: profiling experiments for broad-scale classification

This paper presents a simple prototype sparse detector imaging sensor built using sixteen off-the-shelf, retro-reflective, infrared-sensing elements placed at five-inch intervals in a vertical configuration. Profiling experiments for broad-scale classification of objects, such as humans, humans wearing large backpacks, and humans wearing small backpacks were conducted from data acquired from the sensor. Empirical analysis on models developed using fusion of various classifiers based on a diversity measure shows over ninety percent (90.07%) accuracy (using 10-fold cross validation) in categorizing sensed data into specific classes of interest, such as, humans wearing a large backpack. The results demonstrate that shadow images of sufficient resolution can be captured by the sensor such that broad-scale classification of objects is feasible. The sensor appears to be a low-cost alternative to traditional, high-resolution imaging sensors, and, after industrial packaging, it may be a good candidate for deployment in vast geographic regions in which low-cost, unattended ground sensors with highly accurate classification algorithms are needed.