Research On Non-contact Road Surface Meteorological Condition Recognition Method Using Spectral Analysis

The severe meteorological conditions in winter can cause complex road conditions,such as water,moist,ice,snow cover and other complex states,which will bring out a series of traffic accidents.At present,according to whether the sensor is in contact with the surface to be measured,the road surface meteorological condition recognition methods can be divided into two categories: contact measurement and non-contact measurement.The contact sensor adopts buried installation,which has poor deployment flexibility and high maintenance cost.In contrast,the non-contact method has the characteristics of noninvasive and strong mobility,but the accuracy and reliability of the sensor are easily affected by external conditions.On the basis of analyzing the principle of road surface meteorological condition recognition,this paper studies the corresponding relationship between the road surface condition and its diffuse reflectance spectrum,and proposes a road meteorological condition recognition method based on the road surface reflectance spectrum.Firstly,the optical characteristics of water and ice and the recognition principle of road meteorological conditions are analyzed,and the feasibility of using road reflection spectrum to distinguish road meteorological conditions is determined.Then,the design of the overall scheme of the road meteorological condition recognition system was carried out,and the test platform was built.The method of collecting road surface reflectance spectrum was studied,and the spectral data collection of different road conditions was completed.The influence of road surface texture and detection conditions on the spectral data is also analyzed.Finally,for the spectral data of complex road conditions collected under the best detection conditions,the abnormal spectrum data is eliminated through the abnormal detection algorithm based on Mahalanobis distance;The best preprocessing method is selected through experimental analysis,and the sub-band feature selection and feature extraction method of spectral data of road condition are studied by combining with the characteristics of spectral curve;The model recognition effect of various machine learning algorithms is studied,and through the experimental analysis,the Random Forest algorithm has better recognition effect.It is showed that the optimal number of sub-bands for the sub-band feature selection method is 8,and the number of feature selections for each band is 6;The sub-band feature extraction method which is divided into 8 bands with the slope,area and variance feature extraction can achieve the best effect;The Random Forest algorithm has a recognition accuracy of 96.68% for 12 road conditions,and using the spectrum data of 6 road conditions collected under 4 detection conditions as the test set,the final recognition accuracy is 90.98%,which verifies that this method has strong generalization performance and high recognition accuracy.