Research On Driving Intention Identification Based On Hidden Markov Model

With the rapid pace of modern society and higher demands on the security and comfort of automotive, various advanced technologies are widely used in automotive products, especially electronic technology which already becomes the trend of the times. There are more and more kinds of driver assistant system placed on automotive. To these driver assistant systems, achieving smooth control mode transitions between automated to manual operation is very important. In recent years, more efforts are under way to research and develop X-By-Wire system, which has cancelled the conventional mechanical joint between accessories. In X-By-Wire system, the driving intention is perceived by sensors and is sent to Central Control Unit by wire. The Central Control Unit controls corresponding actuators through control instructions, to realize driving intention. Therefore, the realization of X-By-Wire system and smooth control mode transitions is based on driving intention recognition, whether for a driver assistant system or an X-By-Wire system.This dissertation is based on the project named "Research on Key Technology for Front Wheel Steer-By-Wire Vehicle" and "Research on Control methods and Key Technology for X-By-Wire Vehicle" by National Natural Science Foundation of China. Based on the summary of the domestic and foreign research achievements on driving intention recognition and behaviour prediction, this dissertation aimed at recognizing driving intention at real time with high precision and proposed a model structure with double-layer HMM (Hidden Markov Model) for recognizing driving intention. Firstly, all the Multi-dimension Gauss HMM and Multi-dimension Discrete HMM in the double-layer HMM are trained off line, which represent driving behaviour and driving intention respectively. On the basis of these, the method for driving intention recognition, proposed in this paper, is verified on line with the 29-DOF driving simulator and NI PXI-1042Q Chassis of National Instrument.The main research work is summarized as follows:(1). Proposed a model structure with double-layer HMM for driving intention recognitionAs drivers'driving behaviour has time characters and drivers'driving behaviour sequence under specific driving environment is consistent, a model structure with double-layer HMM for driving intention recognition is built in this dissertation, on the basis of the model structure for layered HMM. The lower-layered Multi-dimension Gauss HMMs correspond to driving behaviour and up-layered Multi-dimension Discrete HMMs represent driving intention, the lower layer of the architecture is connected to the upper layer via its inferential results, thereby an integrated double-layer HMM is obtained. Then combined "HMM toolbox for MATLAB" with MATLAB, off-line train all the Multi-dimension Gauss HMM and Multi-dimension Discrete HMM in the double-layer HMM, which represent driving behaviour and driving intention respectively. All these works lay the foundations for on-line experiment.(2). Selected some working case for diving intention recognition and drew up experimental schemeAs we known, when establishing the HMM-BASED model, parameter determination is the foremost task. In this dissertation, all the Multi-dimension Gauss HMM is trained off-line, by using the driving simulator experimental data. Through the function analysis of each driver assistant systems, active safety systems and X-By-Wire systems, from the "security/comfort" and "single working case/combined working case" point of view, we selected finally'Emergency braking','normal braking','Hill-start','Obstacle avoidance' and'Brake in a turn'these five working cases as the research objects in this dissertation. Such as: Hill-start——main consideration of driving behaviour on accelerator pedal and brake pedal (comfort), (single working case)Obstacle avoidance——main consideration of driving behaviour on steering wheel (security),(single working case)Emergency braking——main consideration of driving behaviour on accelerator pedal and brake pedal (security), (single working case)Brake in a turn——consideration of driving behaviour combined accelerator pedal, brake pedal and steering wheel (security and comfort), (combined working case)With the 29-DOF, stationary driving simulator of Jilin University, drew up experimental scheme for these five selected working case and do corresponding experiment. The entire sensor data acquired in the experiment compose the database for training double-layered HMM.(3).Proposed the method for data preprocessingThe data acquired is pre-processed to make preparations for training. Firstly, change the steering wheel data from radian measure to degree measure; then, a Forward-backward filter is employed to eliminate the noise in sensor data; After data processing including amplification and filtration, the data processed is sorting into three groups which are for pedals, steering wheel and speed respectively, so three data sets can be get; And the data in each data set be cut into several segments, the data segments are sorted according to the short-term driving behaviour to make one segment set only concern one short-term driving behaviour.To data segments for short-term certain driving behaviour, abnormal data segment is discarded with the t-test method with the selected characteristic parameter; and then, we combined with K-means to set the limits for normal behaviour and emergency behaviour on the view of drivers, to verify the correctness of recognized result.(4). On-line driving intention recognitionBecause it is need to identify the double-layer HMM model structure online, so, it is need to acquire manoeuvring signals of driver and vehicle state online, and real-time recognize driving intention on the basis of "HMM toolbox for MATLAB". With the analysis of the advantages and disadvantages of MATLAB and LabVIEW, introduce the M-files of MATLAB into LabVIEW program by using of the'MATLAB Script' control which is provided by LabVIEW. In the mixed program of MATLAB and LabVIEW, MATLAB is charge of running programs which are in the "HMM toolbox for MATLAB" And LabVIEW is responsible for designing graphical user interface, controlling hardware, acquiring sensor data, data analysis and processing, data storage, operational control and network communication.Because that HMM theory is based on the Expectation Maximization algorithm and driving intention MDHMM in upper layer is not for all working case, we set the limits of likelihood for four driving intentions, by means of many online tests. Just over the limit of likelihood, homologous driving intention under given observation sequence cases can be confirmed.The on-line recognition results show that the method proposed in this dissertation has high precision and real-time characteristic.