Pedestrian detection based on maximally stable extremal regions

This paper presents a new approach to generate hypotheses about the presence of pedestrians in an infrared image. Information about maximally stable extremal regions is used to locate the warmest regions on the image, which are considered to be potential human heads. To capture the complete human body, these regions are scaled based on the range data of a lidar sensor. Closely related regions are merged into one bigger region to avoid the segmentation which arises from the heterogeneous heating emission of a dressed human. Additionally, the area and perimeter of each potential pedestrian are examined to discard artificial objects. The optimal decision measure is sought so that all pedestrians are extracted from a scene. All remaining hypotheses should be further processed with a statistical classifier.