Abstract: Pedestrian detection is an important research field in computer vision.The key of pedestrian detection is feature selection.Most of pedestrian features are based on Papageorgiou’s rectangle features.In most circs，the pedestrian in video is small and faintness.The shortcoming of rectangle features is that they are not good in describing edge of pedestrian.In this paper，four new features called triangle features are presented and trained by an adaboost algorithm to get a representation of pedestrian.Unlike rectangle features，calculating triangle features is not easy. An algorithm to calculate triangle features is proposed in this paper.Compared with rectangle and asymmetry rectangle features，triangle features are good in describing edge of pedestrian.A set of experiments indicate that triangle features are better in describing local characteristics of people’s appearance and designing pedestrian detector.