The laser scanner provides non-contact, high accuracy and high speed probing and has found many applications in manufacturing product inspection. Considerable effort has been put into extracting geometric features from point clouds. However, the research results for feature extraction from edge detection and computer vision do not meet the metrology requirement of high accuracy. Our test results have shown that features extracted from a point cloud obtained from laser scanning have poor accuracy compared to the result measured with a touch trigger probe, even with very careful qualification of the laser scanner. Generally, a circle extracted from a point cloud will be larger and the centre will be shifted compared with the result from a touch trigger probe. This paper investigates the causes of poor accuracy in laser scanning for geometric features and develops an algorithm to solve these problems. The test results show the algorithm developed works well and inspection accuracy and calculation speed have significantly improved.