A multi-stage image segmentation framework for human detection in mid wave infra-red (MWIR) imagery

Abstract

Detecting Human objects in Thermal imagery is commonly realized in two steps. The first step is segmentation; it identifies Region of Interest (ROI) in a given image that likely to contain a human target. The second step is Classification; the identified ROI is verified for human objects. Accurate segmentation step can significantly boost classification accuracy. We present a multi-stage image segmentation framework for surveillance applications, to detect human targets present at widely varying ranges in thermal imagery. A temporal median filter is utilized to estimate the background frame from previously sampled frames. Using the estimated background image, clutters are suppressed with K-L Transformation. A box filter is applied on clutter suppressed images for identifying the Region of Interest (ROI) objects. Finally the identified ROIs are subjected to two level segmentation process for accurately extracting the silhouette. The performance of proposed method is analysed on three thermal Infra-red video sequences containing targets in the range of 500m, 1 km and 1.5 km. This way of demonstrating performance with range is first of its kind in human target detection literature. Experimental results demonstrate that proposed segmentation approach is able to extract human silhouette without loss of significant shape information even for targets at far ranges and with background occlusions.