Data-Driven Skin Detection in Cluttered Search and Rescue Environments

Abstract

Locating human victims in cluttered urban search and rescue (USAR) environments is still a challenge. In this paper, we present an approach to generate skin objectness windows to assist human rescuers. We introduce the term skin objectness to denote the task of extracting windows in the scene with a high probability of skin presence for locating victims. Unlike naive skin segmentation approaches, the presented algorithm accounts for both color and spatial information to extract regions of interest and at the same time, rejects the background clutter. We use temporal information of the video sequence to make the skin objectness windows more reliable. To selectively boost skin regions, the RGB skin pixels are transformed to Gabor space to generate a transformation matrix. The matrix is used to generate skin affinity maps removing the unwanted background clutter. Further, the Bayesian inference and temporal cues from previous frames are used for detecting skin objectness windows. It has real-time applications in image retrieval, action classification, virtual reality, etc. The experimental analysis demonstrates quantitative and qualitative results on a disaster victim dataset showing plausibility and efficiency of the proposed method in cluttered environments.