A Data Annotation and Recognition Method Based on Zero Statistical Hypothesis Test and Multi Variable Binary Classification Theory

Abstract

Based on typical Chinese natural driving data, from natural driving scenario data collection to scenario automatic labeling and classification, this paper proposed a specific scenario automatic labeling and classification method by using statistical tools and machine learning methods. The front vehicle cut-in data of more than 4000 typical road scenarios in China are collected and extracted, and the parametric statistics and analysis are carried out for the relevant 6 variables. Considering the statistical uncertainty of the variables, the statistical exclusion curve of “normal front vehicle cut-in scenario” is calculated by using the hypothesis test method based on the principle of mathematical statistics, by comparing the distribution curve of any event with the statistical exclusion curve, the annotation of the front vehicle entry scenario data is realized. At the same time, using the positive and negative sample classification method of machine learning based on bagging decision tree classifier, the integrated learning classification method based on boosting decision tree, and the depth learning method based on improved resnet-18 convolution Network + LSTM recurrent neural network, the multi-variable binary classifiers are trained respectively to realize the classification task of the front vehicle cut-in scenario. Furthermore, comparing the three classification methods, the test results on the verification show that the BDT classifier has the best result, effectively realizes the classification tasks of “dangerous front vehicle cut-in scenario” and “normal front vehicle cut-in scenario”, and this technical tool chain can be reused in the fine-grained classification of other driving scenes in the future