Natural corrosion of steel bars in service structures differs from experimental artificial accelerated corrosion. For investigating the natural corrosion characteristics of steel bars in practical engineering， a batch of corroded steel bar specimens was acquired from reinforced concrete bridge decks that have been in service for more than 50 years. Precise residual surface information of the corroded steel bars was obtained by three-dimensional laser scanning， and corrosion all along the bars was described by residual cross-sectional areas. The effects of corrosion degree and bar length on longitudinal corrosion non-uniformity were studied. The probability distribution characteristics of longitudinal stochastic corrosion were studied by statistical analysis， and the corresponding probability distribution models were proposed. Results show that naturally corroded steel bars have various typical corrosion morphology. The corrosion zone and corrosion degree are random along the bars. The maximum cross-section loss ratio and variance of residual cross-sectional areas increase as the average corrosion degree increases， which indicates that the more severe the corrosion is， the more uneven the corrosion is. Bar length influences corrosion non-uniformity significantly. In the case of the same source and average corrosion degree， the non-uniformity increases with the increase of bar length. The longitudinal stochastic corrosion of steel bars has a statistical law. The probability distribution of residual cross-sectional areas of steel bars with different corrosion degrees can be accurately evaluated by the multimodal normal distribution model. In comparison， the Weibull distribution model and the unimodal Normal distribution model have a slightly poorer fitness， and are simple in form and easy to apply， which are suitable for steel bars with an average corrosion degree of less than 10%.