Driving Pattern Recognition Based On Maximum Mean Discrepancy Transfer Learning

Transfer learning is a very important task in machine learning,which has attracted the attention of many scholars.In practical applications,such as image recognition,text categorization,autopilot,etc.,it is often very difficult and costly to obtain a large amount of labeled data for training.What we can get is often a small amount of labeled data or a large number of labeled data in other related fields.With the continuous development of autonomous driving technology,the problem of driving pattern recognition has become an increasingly important issue.It is closely related to the perception,decision making and control of autonomous vehicles.This paper use transfer learning to solve the problem of lack of labeled data and huge data distribution differences in the automatic driving pattern recognition problem.Maximum Mean Discrepancy(MMD)has been used to evaluate and describe the distribution similarity between source and target domain in transfer learning problems,but they are often only for global data and do not focus on intra-class compactness and inter-class separability.This paper focus on the problem of driving pattern recognition based on the maximum mean discrepancy transfer learning.The main research work and contributions are as follows:(1)This paper proposes a robust driving pattern recognition method based on intraclass compactness and majority voting classifier.This method mainly considers the important influence of intra-class compactness on transfer learning,and uses the majority voting method to concentrate the group wisdom of multiple classifiers to achieve more robust classification results.The experimental results on the actual real dataset further verify the effectiveness of the proposed robust driving pattern recognition method.(2)This paper proposes a method to combine intra-class compactness and interclass separability for driving pattern recognition.The combination of Intra-class Maximum Mean Discrepancy and Inter-class Maximum Mean Discrepancy not only increases the similarity of data distribution within the same class of the source domain and the target domain,but also increases the difference between the different classes of the source domain and the target domain.The experimental results of this method on the parking lot dataset further validate the validity.Comparisons with other algorithms show that the proposed metric is valid.Therefore,this paper proposes a model combining the two measures of Intra-class Maximum Mean Discrepancy and Inter-class Maximum Mean Discrepancy.This model can simultaneously enhance the compactness of data within a class and the separability of data between classes.Such a transfer learning method can better transfer the knowledge of the source domain to the target domain for better classification.