System Design of Driving Behavior Recognition Based on Semi-supervised Learning

Abstract

Driving Behavior Recognition (DBR) has always been a key problem in the field of vehicle driving and traffic safety. However, there are some problems in previous studies that only the coarse-grained features are extracted and a large number of unlabeled samples consume human and material resources to label them. In this paper, we propose a driving behavior recognition system based on semi-supervised learning to solve these problems. First, wavelet decomposition is used to process the sensor signal, giving a decomposition of sensor signal into a set of approximate and detailed coefficients. Then, we extract 300 features from the decomposed signal to capture the patterns of driving behaviors, which contains fine-grained features especially. And 34 features are selected by a feature selection algorithm based on random forest algorithm (RF). Finally, an improved semi-supervised algorithm is proposed to optimize the strategy of selecting the unlabeled samples. And the experiment results show that fine-grained features are also effective and important relatively. And the improved algorithm has better classification performance compared to the previous algorithm using the dataset with different labeling rate.