Improved machine learning approaches for individualized human assistance, supervision, and behavior prediction

Abstract

In recent years, research and development of predicting driving behaviors play an important role in the development of Advanced Driver Assistance Systems (ADAS). For this reason, many machine learning approaches have been developed and applied in this field. Due to the advantages of Hidden Markov Model (HMM) in dealing with time series data as well as state transition descriptions, the HMM seems to be a suitable algorithm in driving behavior prediction. Therefore, one of the aims of this thesis is to analyze the current state of various driving behavior models and related HMM-based algorithms. Except for using a single HMM to establish a driving behavior model, two design ideas (HMM-derived or HMM-combined approaches) can be concluded from the existing research to improve the HMM performance. Based on the two design ideas two newly developed approaches named Fuzzy Logic-based Hidden Markov Models (FL-HMM) and Multi-Layer HMM (ML-HMM) are designed based on HMM-combined and HMM-derived approach in this thesis. To determine and predict drivers behaviors in the future, the main idea is to learn the driver's historical behaviors. For this reason a model has to be established and trained first. To improve the training process, in this thesis a strategy is developed for higher reliability in terms of accuracy, detection rate, and false alarm rate. The strategy is named full scale training loop and can be used to optimize both model structure and model training. Based on the proposed approach, seven algorithms including five conventional algorithms (HMM, Support Vector Machines (SVM), Artificial Neural Networks (ANN), Convolutional Neural Network (CNN), Random Forest (RF)) and two new approaches (FL-HMM, ML-HMM) are used as examples to identify driving behaviors. To improve the prediction performance of the related models, design parameters, which are unknown and need to be set manually before training, are modified. Using the proposed training procedure the most suitable design parameters can be determined automatically to optimize the performance of the algorithms. In this thesis, design parameters are divided into two categories: hyperparameters and prefilter. The focus is to demonstrate the ability of the proposed approach to improve the prediction performance of different algorithms, and to discuss the effects of hyperparameters and prefilters. Based on the data achieved from 17 drivers, four different models are designed for each algorithm to validate the effectiveness of using hyperparameters and prefilters. The finally obtained results show that the prediction performance can be improved using the proposed optimized training procedure.