Night Pedestrian Detection Method Based on Image Fusion

Abstract

Improving pedestrian detection accuracy under conditions of insufficient light at night is a priority for automatic driving. The clarity of conventional visible images at night cannot be as good as that during the daytime. In contrast, the images captured by infrared cameras are almost unaffected by ambient light changes, but there are also defects of not apparent texture features. Therefore, the pedestrian characteristics can be enhanced by combining the thermal and the visible images, which are not affected by the ambient heat. There have been many types of research on pixel-level image fusion in medical diagnosis, military enhancement and navigation, and it has an excellent application value. In this paper, pixel-level fusion method is applied to the fusion of thermal and visible images, and tests on the pedestrian Dataset CVC-14 acquired in a real-world environment. The YOLOV5 method, which has significant advantages in recognition speed and application flexibility, was used for detection. The experimental results showed that the average recognition accuracy and iteration speed was superior to that of the source image. The performance of maximum fusion, principal component analysis (PCA) fusion based on thermal imaging, and weighted average fusion was outstanding. The method proposed in this paper is not complicated and widely applied, which can be improved by other researchers based on this direction.