Research Of Road Surface Condition Recognition Key Techniques Based On Machine Vision

The transportation meteorological observation system (RWIS) studies have been become an emerging research application domain. They gradually receive great attention at home and abroad. In recent years, along with the increasing intelligence transportation system (ITS) development, real-time monitor, accuracy and robustness on the various RWIS measurement elements face the new requests. Among them, as one of impact road safety key factors, road condition monitor and identification can not only safeguard the safe driving, but raise the overall transportation system's operational efficiency. Moreover, it enhances ITS intelligence management level. As it is difficult for the special working environment to model the road conditions, the problem of low recognition rate on road condition sensors has always existed and limited its research and application development. In China, although many stations and major highway junctions have been installed video surveillance equipment, road conditions and other important information as security for the highway still needed to get through the human-computer interaction. Some road status information even rely on manual monitoring of access. In this situation, the meteorological service can hardly satisfy intelligent transportation systems (ITS) requiring the use of accurate weather requirements. It is important for automatic road condition recognition to reduce traffic accident, enhances ITS decision-making effectiveness.This article intends to meet highway safety management actual need and meteorological technical instrument demand. It has conducted the deep research regarding the key technologies which have severely affected the correct recognition's rate for road condition. These include the road image calibration problems because of the light source, the condition characteristic parameter extraction problems and the pattern recognition methods used for different characteristic parameters. And these methods were used to conduct training and test the images acquired from the physical highway environment. The correction algorithms based on the monitoring color constancy were studied and implemented. The color and texture feature extraction methods and their identification methods of road surface image were studied in-depth. As the result, these have greatly reduced the light source stability requirements, improved road condition recognition method robustness, enhanced road condition identification accuracy. Acquired advances were involved in this work:1. The road condition image sensor was designed. To conduct research in road condition recognition methods, a CCD camera was selected as the front-end access to means of images. Based on the Windows XP operating system, MATLAB 7.6 and Visual Basic 6.0 design language were adapted to develop the road condition image sensor and its software. At present, it can recognize four kinds of road states.2. Based on the surveillance color constant, one kind newly intensity illumination correction method under low light conditions was proposed. The illumination adjustment problem caused by light color temperature change in road images was solved. According to the human vision contrast resolution's limit rule, the parabola function was used to take brightness gain function. Image brightness values were mapped into the best recognition scope defined by experiments. As a result, stability request for the recognition algorithm in environmental illumination change was obviously reduced. The recognition algorithm robustness hereafter was guaranteed effectively.3. According to the road surface reflection image characteristic, for the first time, the color feature was extracted to segment the road surface reflection image. To segment color road surface images, the segment effect was satisfying using B/R of RGB color system. Following, was that of the H image of the HIS system. The segment precision was obviously higher than the traditional one of using texture gradation matrix.4. The wavelet packet texture features from road images were analyzed and extracted. The road surface reflection image signals mainly concentrate in the intermediate frequency and high frequency band. Thus, wavelet packet analysis method was employed and the texture details were retained effectively.5. A new road condition recognition method based on wavelet-HMM was put forward. Through wavelet packet decomposition, the fast training and the likelihood algorithms were used to extract image wavelet packet sub-band energy as texture features. And they were input into HMM to be trained to take HMM parameters so that the road condition classifiers were established and road surface condition recognition was come true. In the method, the wavelet coefficient relevance was considered. And the inaccurate question in describing wavelet coefficient distribution with the standard gaussian distribution was corrected.6. On the basis of extracting the information in two kind of colored models to form feature vectors, a improved BP neural network recognition algorithm was proposed. It combined with both additional momentum method and the auto-adapted study rate. Based on this, a misalignment recognition model was established and applied in the road condition recognition experiments. The experimental result indicated that this model can distinguish the road surface meteorology condition accurately.In the simulation experiments the both of road recognition methods all successfully recognized the four kinds of road condition.