Research On Some Key Techniques Of Road Object Recognition Based On Machine Learning

Along with the popularization of vehicles, traffic accidents have become one of the biggest social problems all over the world. Especially in China, the losses, which are caused by traffic accidents in both lives and economy, are startling. It is worth the whistle that about 85% of these accidents are due to human factors worldwide and even 95% in China. Therefore, it has become the key research direction in ITS (Intelligent Transportation System) field to promote the vehicle active security through improving object recognition performance and providing drivers with more alert and assistance information using the techniques of sensors such as video and radar sensors. On-road object recognition based on video sensors has become one of the focuses due to the low cost and the wide vision scene. At present, most methods of road object recognition based on vision follow two steps:hypothesis generation (HG) and hypothesis verification (HV). HG generates ROI (Regions of Interest) which include candidate objects; HV verifies the existence of object on ROI. Machine learning is one of the main methods of HV and has become an important research topic of vision-based road object recognition for its potential and availability. Currently, many machine learning based methods have been proposed for on-road object recognition. After systematically analyzing current problems of machine learning based object recognition, this thesis focuses on the following key techniques of machine learning:feature extraction and reduction, parameter selection of SVM (Support Vector Machine), AdaBoost algorithm, samples automatic preparation and the classification problem caused by imbalanced training samples. These key techniques are studied systematically.Firstly, considering the problem that the current Haar wavelet feature extraction methods based on signed coefficients on the grayscale image are sensitive to the illumination and poor anti-noise, a Haar-like feature extraction method is proposed. Compare to the current typical feature extraction methods, this approach not only has better performance, but can be calculated rapidly. Furthermore, the dimension of feature vector obtained with the current feature extraction methods is usually too high, so a feature reduction approach is proposed. On one hand, this method can reduce the dimension and can avoid the difficulties during classification calculation, on the other hand, it can remove the redundant information and noise of feature vector and improve the generalization of classification.Secondly, because the current parameter selection of RBF-SVM is time consuming, an improved parameter selection method is provided, by which the time consumption can be reduced in the period of parameters selection and training, while the detection performance is closed to that of the traditional cross-validation method.Thirdly, considering the time consumption of AdaBoost classifier preparation, an improved construction method of weak classifier of AdaBoost is proposed, which can reduce time consumption of the training process of AdaBoost classifier, moreover, the proposed self-adaptive threshold setting method can overcome the problem that the traditional threshold setting method can't reflect the distribution of training samples. Considering the time consuming of the training process of AdaBoost classifier itself, especially the great scale training samples, an increasing learning approach is proposed to effectively improve AdaBoost classifier's ability and learning efficiency.Fourthly, on the analysis that current sample preparation is strongly dependent to manual operation, an automotive sample preparation method is proposed, which aims to deal with the problems of serious subjectivity and burdensome task when preparing the samples. In addition, since the classifier will bias to multi-class samples when the samples is imbalanced, a SVM ensemble based method is proposed to solve the training of imbalanced samples, this method reconstructs the training samples by using interval sampling method. The proposed method can overcome the problems of losing classification information, time consumption and unstable detection performance caused by random sampling.Finally, the achieved research results are applied to the rear-object (including vehicle and motorcycle) detection system based on monocular vision. The conclusions under different time, different weather and different background conditions are drawn as follows:the test results under twilight and bad weather (including rain, snow and fog) scenes demonstrate the effectiveness of the presented algorithms, and the test results under sunny and cloudy weather at daytime prove the effectiveness and availability.