摘要：The visual measurement restricts the navigationaccuracy of the vision-aided integrated navigation system.Thus, how to obtain the visual measurement quickly andaccurately which involves the feature extraction becomes akey focus. Among the various feature extraction methods, themost commonly used feature extraction methods are the scaleinvariant feature transform (SIFT), the speeded up robustfeatures (SURF) and the features from accelerated segment test(FAST). The performance evaluation is beneficial to choosingappropriate feature extraction methods for visual measurements.Although a great many of studies on their performanceevaluation exist, there is lack of performance comparisonamong the abovementioned three feature extraction methods.Therefore, researching on the evaluation of SIFT, SURF andFAST is of great importance, which is the main objective ofthis manuscript. In this paper, the theoretical principles ofthese three methods were firstly overviewed. And then theirperformance was compared and analyzed from three aspects:the computing time, the capability of extracting features andtheir invariances. In order to make the comparative analysissystematically, the sequences of the image transformations usedin this paper were carried on rotation, scale, blur, compressionand illumination, respectively. The experimental results showedthat among the three methods, the FAST method was the fastestone and the SIFT method possessed the strongest extractioncapability. The rotation, scale and compression invariances withthe SIFT method were all superior to the ones with the othertwo methods. For the blur invariance, the SIFT and SURFmethods had similar performance which was better than theone of the FAST method. Besides, the illumination invariancewith the FAST was not as good as with the other methods.
关键词：scale invariant feature transform (SIFT);speeded up robust features (SURF); features from accelerated;segment test (FAST); feature extraction method; invariance;performance evaluation
