Research On Road Recognition Algorithm In Complex Environment Based On Incremental Learning

Road recognition technology based on machine vision is one of the core part of autonomous land vehicle. Although many scholars spent a lot of time having further study on the road recognition technology, but there are still many problems restricting the further development of road recognition technology. Considering the reason, there are the complexity and diversity of the actual road environment, such as, shadow, water stains and leaves mulch, so it greatly increasing the difficulty of road recognition. In addition to, in order to enhance the generalization performance of the decision function and improve the right accuracy of road recognition in the complex environment, it often needs a lot of different training samples to train the classifier. So if it process the traditional batch learning style, there are spending a lot of computing resources and it is difficult to adapt to the actual application requirement.According to the above problem, the core research on incremental learning and road recognition under the complex environment in this paper, the main work is as follows:(1) On the basis research of twin support vector machine and drawing lessons from incremental learning algorithm of traditional support vector machine, this paper proposed a new algorithm-an incremental learning algorithm based on twin support vector machine. It divides the training sets into historical training samples and new training samples. When new training samples joins the training sets, the classification decision function has to be updated. Keeping the important samples of the historical training samples and selecting the core samples of the new training samples, so it create an effective dynamic training sets. Under the premise of twin support vector machine prediction precision, it processes incremental learning and reduces computing scale greatly.(2) On the further research of twin support vector machine, it expand classification problem to regression problem and proposed a new algorithm-an incremental learning algorithm based on twin support vector regression. Different from creating dynamic training sets, it makes full use of old computing information instead of learning all training sets, so it greatly simplifies the calculation of inverse matrix and improves the execution efficiency. Experimental results for artificial datasets, time series and road datasets show that this algorithm has remarkable improvement of generalization performance with short training time, and get a general road recognition.(3) This paper proposed a new road recognition algorithm under complex environment based on K-means feature. Firstly, using SLIC (Simple Linear Iterative Clustering) algorithm to segment original road images, they were divided into irregular image patches in which the pixels are homogenous and at the same illumination level. Secondly, it extracted K-dimension features from super pixels by the means of K-means clustering. Then, the training set was made up of K-dimension features. Thirdly, it trains twin support vector machine through incremental learning style and gets the classification decision function. Finally, using same method to extract features from road images, then the classification decision function classified the road and not road region. Simulation results show that this algorithm to some extent solve the time-consuming training problem and low road recognition accuracy under complex environment.