Study On Road Area Detection Based On Machine Learning Methods

Road area recognition is one of the supporting technologies for autonomous navigation system and autonomous driving. It relates to computer vision, pattern recognition and many other research areas and the solving of this problem has important significance in both research and practice.In recent years, road area recognition system based on visual information received wide attention because it is easy to be equipped and costs little. However, due to the weak adaptability for complex road scenes and illumination change, road detection systems based on visual information may perform poorly in robustness and generalization.Our research is based on road area detection using visual information. In the research, deep learning methods are introduced in road scene feature extraction sector and together with the supervised learning framework we will be able to achieve the goal of detecting the road area. The main contents and innovations are as follows:1. Different from the commonly used feature extraction methods like color and texture, DBN network based on deep learning for feature extraction is introduced which can to some extent avoid subjectivity and improve the stability and robustness of the detection algorithm.2. A road area detection method is proposed based on SVM posterior probability and Graph Cuts. By making full use of the posterior probability information provided by the classifier, our method can provide more stable and reliable results.After researching on the updating strategy of the training data, experiments are done on multiple datasets composed of consecutive frames. Experimental results show that the algorithm performs well in stability and robustness and satisfactory detection results can be obtained in a variety of complex scenes.