서강대학교

초록/요약 도움말

This thesis targets the detection of pedestrians in far-infrared images that can contribute to the reduction of critical traffic accidents, especially at nighttime. The primary goal for detection systems is to develop algorithms suitable for far-infrared images, which have distinct characteristics that differ from visible spectrum images. To achieve this goal, our thesis focuses on the development of a robust candidate generation method and a feature extraction method that specializes in far-infrared images by utilizing the unique characteristics of these types of images. This thesis makes three main contributions. First, we propose regions of interest generation method that is based on image segmentation. The idea for this approach comes from the statistics of far-infrared images; the approach has the great advantage of the ability to extract intrinsic images of pedestrians. As a result, the proposed method generates a small number of candidates at an acceptable miss rate, and the generated candidates provide advantages for classification because the pedestrians are well-arranged within a bounding box. Second, we propose an intensity-based feature extraction method. The basic idea for this method is that local gradients can be estimated using intensity differences between neighboring pixels because far-infrared images are characterized by monotonic gray-level changes. This feature has advantages for real-time systems because it involves simple calculations with satisfactory performance. Last, we show performance improvements in using pose-specific and multiple feature-based classifiers. Local classifiers trained on two kinds of poses and two complementary features are fused to obtain better performance. Experiments show that the proposed classifiers improve the classification performance of single feature-based classifiers.