Human visual attention based saliency model is an effective way for the active perception of important information in image, and it can play an important role for exploring large-scale perceptual information organization in the early stage of visual cognitive process. However, nighttime images usually have low signal-to-noise ratio and low contrast properties. Conventional salient object detection models may face great challenges in this scenario, such as noise influence, and weak texture blur. This paper proposes a region covariance and global search based salient object detection model for nighttime images. The input image is firstly divided into superpixel regions to estimate their covariance. Then, covariance feature based local saliency and global search based image saliency can be calculated respectively. Finally, a graph-based diffusion process is performed to refine the saliency maps. Extensive experiments have been conducted to evaluate the performance of the proposed method against eleven state-of-the-art models on five benchmark datasets and a nighttime image dataset.