Automatic Curve Detection Based On Machine Learning

Curve detection is an important research aspect in the field of intelligent transportation. The traditional road models based on curve detection methods require establishing a curve model to fit lane lines. But one single curve model is not suitable for different shapes of curves, further more, curve model is complex and its calculation is large, so it often can’t meet the requirements of accuracy and real-time performance for curve detection. This thesis uses machine learning methods to realize automatic curve detection under different weather conditions. Machine learning methods obtain the rule through learning from the existing samples, the rule is then employed to predict unknown samples. Compared with traditional methods, this method uses implicit rules of curves other than curve model to detect unknown curves.In this thesis, some machine learning algorithms, such as Support vector machines (SVM) and AdaBoost, and PHOG feature are employed to detect curves under different weather conditions. SVM is based on statistical theory and the principle of structural risk minimization, which uses kernel function in the feature space to solve nonlinear problems. AdaBoost which is on the basis of boosting algorithm, combines the output of weak learners into a weighted sum as the final output of the boosted classifier. As a spatial shape descriptor, PHOG can represent the overall shape, the local shape, and their spatial relationship, and has some capabilities of anti-rotation and anti-noise.The image database constructed in this thesis contains left and right curves under different weather condition, in which the left and right curves are subdivided into four classes, respectively. SVM and AdaBoost algorithm are employed to detect curves with the image database. Experimental results show that the detected accuracy of SVM reaches above90%, while AdaBoost algorithm reaches above77.5%, and SVM has the better robustness than AdaBoost algorithm. As the weather (rain, snow and fog) gets worse, the accuracy is decreasing gradually, and the fogy weather has more impact on the performance degrade than the rainy and snowy weather. For this problem, the thesis employs the method of dark channel prior to remove the fog. Experimental result shows that the accuracy has improved.