As the pedestrian detection with vehicle-mounted far-infrared monocular sensor using machine learning is usually poor in real-time performance and precision ， a head-histogram of oriented gradient-support vector machine （ Head-HOG-SVM ） approach based on edge segmentation is proposed. The weighted Sobel operator is adopted to enhance the vertical edges of pedestrians in the regions of interest （ ROIs ） . Several pedestrian detec-tion methods are selected according to the pedestrian appearance in different distance. A head feature is used to detect pedestrians at near and middle distance to improve the real-time performance of the system ， and a HOG-SVM classifier cascading with head recognition is used to detect blurred pedestrians at far distance.Experimental results on the several videos captured from suburb scenes show that ， in comparison with the HOG-SVM classifier based on dual threshold segmentation ， the precision and detection rate of the proposed method are respectively increased by 33% and 200%.