In this paper, we present a robust method for nonisotropic point light source calibration through feature points selection. By analyzing the relationship between the observed surface and its image intensity under near-field lighting, the feature points selection method is first developed to effectively address the noisy observations and improve calibration robustness. Afterward, to enhance efficiency and accuracy of the calibration, a cost function of l p-norm is established based on the above relationship, and an improved Newton method-based iteration process is applied to calculate the light source parameters. The simulations demonstrate that the proposed method is capable of achieving robust calibration results with the estimation error less than 2.7 mm and 0.8°, even though the image intensities are corrupted by Gaussian white noise with standard deviation up to 0.4. The experimental validation is performed using a self-designed photometric stereo system, where the calibration of point light sources is conducted, and measurements are taken on a standard sphere and compressor blade based on the obtained calibration results, which demonstrates the effectiveness of what we believe to be a new method.