The invention discloses a label extracting method based on perception data. Object data includes numerical value data and text data. During label extraction, the two types of data are processed separately, and labels of the two types of data are related through probability statistics after being extracted. During numerical value feature label extraction, the mass center is selected through clustering to serve as a final label of the type, and the clustering quality is directly related to the effect of feature label extraction. The labels represent the most outstanding feature point of a certain type, that is, the difference of all examples in the type is the minimum on semantics. Most of clustering algorithms abide by the principle, but actually the numerical value feature labels do not need to represent feature semantics of the cluster. The sequence feature label extraction process is mainly divided into clustering and mass center selecting. Similarity matching is conducted on the introduced perception data and the labels in a label library to obtain corresponding numerical value labels. The numerical value labels are related to a text label library to obtain corresponding text labels, and the text labels are returned.