This paper presents a part-based human detection method that is invariant to variations in the view of the human and partial occlusion by other objects. First, to address the view variance, parts are extracted from three views: frontal-rear, left profile, and right profile. Then a random set of rectangular parts are extracted from the upper, middle, and lower body as the distribution of Gaussian. Second, an individual part classifier is constructed using random forests across all parts extracted from the three views. From the part locations of each view, part vectors (PVs) are generated and part bases (PB) are also formalized by clustering PVs with their weights of each PB. For testing, a PV for the frontal-rear view is estimated using trained part detectors and is then applied to the trained PB for each view class. Then the distance is computed between the PB and PVs. After applying the same process to the other two views, the final human and its view having the minimum score are selected. The proposed method is applied to pedestrian datasets and its detection precision is, on average, 0.14 higher than related methods, while achieving a faster or comparable processing time with an average of 1.85 s per image.