A concurrent neural network approach to pedestrian detection in thermal imagery

Abstract

The paper presents an original approach for pedestrian detection in thermal imagery using the neural network classifier called Concurrent Self-Organizing Maps (CSOM), previously introduced by first author; it represents a winner-takes-all collection of neural modules. The considered algorithm has the following main stages: (a) detection of the regions of interest (ROI); (b) feature selection using the Histogram of Oriented Gradients (HOG); (c) classification using a CSOM classifier with several neural modules for each class; (d) decision fusion of the SOM modules into the two final classes: pedestrians and non-pedestrians. For training and testing the proposed algorithm, we have used the OTCBVS-OSU Thermal Pedestrian Database provided by the Ohio State University. One obtains the False Positive Error Rate (FPER) of 1.79%, the False Negative Error Rate (FNER) of 0.49% and the Total Success Rate (TSR) of 98.48 %.