Edge information is often the key to the detection of objects. Traditional edge detection algorithms calculate gradient omnidirectionally, which usually results in calculation of redundancy. Inspired by Weber local descriptor, this study proposes a saliency texture structure descriptor that simulates divergent and significant characteristics of human eyes observing things. First of all, it calculates the sum of relative differences between intensity of a center pixel and those of its neighborhood pixels, and divide the sum by the center pixel's intensity to get its local saliency factor. Then, it extracts its local texture structure though a divergent gray level co-occurrence matrix. At last, it constructs a two dimensional histogram as the feature vectors by combining saliency factor and texture structure. Experimental results show that the saliency texture structure descriptor has the ability of good edge detection and powerful structural expression, and is robust to noise and light and shadow changes. When used in pedestrian detection, the saliency texture structure descriptor gets much higher detection rate than other local descriptors such as CENTRIST and HOG. This descriptor can find its high application value in vehicle active safety system.