Research On Identification Of Driver Intention Aiming For Pedestrian Collision Warning

The vehicles have become the mainstay of the transportation from the 19th century to the 21st century. At the same time, the development of the automobile industry also caused serious environmental problems and traffic safety problems. People also transferred the study about the traffic safety problems from improving the vehicle safety level simply to considering the factors such as vehicles, drivers and the environment, to improve road traffic safety. In traffic safety problems, the main traffic accidents are the collisions between vehicles and pedestrians. And the pedestrians are the largest vulnerable groups. Therefore, how to improve the active safety performance of the vehicles and protecting the safety of pedestrians effectively have been more and more valued. Supported by the project of National Natural Science Foundation of China (61104165) and the Fundamental Research Funds for the Central Universities (DUT13JS02), the thesis carries out the research on identification of driver intention aiming for pedestrian collision warning.Based on the pedestrian detection in front of the vehicle, four kinds of driver intention are determined in this thesis, acceleration, deceleration, normal steering and emergency steering. By analyzing the mechanism of driving intention and the statistical pattern recognition theory, the Hidden Markov Model (HMM) is used to identify driver intention in this thesis. According to the four kinds of driver intention, the sensor data needed for the experiment are collected by the driving simulator, and the experimental data are need to be processed. In this thesis, the t-test is used to eliminate the outliers of the sensor data, and the improved K-means algorithm to determine the threshold of the normal steering and emergency steering.Trough the introduction of the Baum-Welch algorithm and the forward algorithm of the HMM, this thesis uses MATLAB in combination with Baum-Welch algorithm and the forward algorithm to write m file to identify driver intention. First, the Baum-Welch algorithm is used for the offline training of HMM. Because of the different length of the observed sequence between different models, and in order to get the more accurate model in the experiment, the Baum-Welch algorithm is improved in this thesis. The HMM parameter to characterize the driving intention is got by training. Finally, the observed sequence after processing is used to identify driver intention online. The observed sequence is input to the completed HMM, and the forward algorithm is used to get the likelihood between the observed sequence and different models. The model with the maximum likelihood is the driving intention HMM. Based on the pedestrian detection result in front of the vehicle by the team, after the identification of driver intention in this thesis, the pedestrian collision warning mechanism is determined. For the wrong driver action, such as the error stepping on the accelerator pedal, the thesis formulates the mechanism to warn the driver and pedestrian at the same time, to protect the pedestrian effectively.