Seeing Signs of Danger: Attention-Accelerated Hazmat Label Detection

摘要

Rescue robots and similar vehicles must recognize various visual objects. Some are of particular interest and must be reliably recognized, for example, hazard signs. Hazmat labels and other intentionally placed signs of danger are typically attached to walls, containers, or vehicles, in locations where they attract attention. These backgrounds typically are of relatively simple structure (though not guaranteed to be plain) while the labels have saturated colors and high contrasts. We provide a new dataset that contains such images and a novel hazmat detection method. It includes an attentional preselection, which exploits the salient design and placement of the labels to locate them, followed by a SIFT-based classification that determines the concrete label type. The results show substantial speed improvements and accuracy gains over the traditional method without an attention stage.