Study On Driver Fatigue States Identification Based On Improved HMM

Driver fatigue plays an important role in causing road traffic accidents,especially serious traffic accidents on highway.The number of accidents will be reduced greatly if driver fatigue can be detected and driver can be warned timely.So study on driver fatigue states has important theoretical research significance and practical use.In recent years,detecting driver fatigue based on machine vision is becoming research hotspots because of its excellent advantages.However,the generation of driver fatigue is dynamic and gradual and recognizing driver fatigue states based on machine vision deduces indirectly driver fatigue states according to observation information.So,to detect driver fatigue efficiently must take above factors into consideration.The double randomness of Hidden Markov Model can reflect reasonably the dynamic relationship between fatigue states and observation variables.So this paper establishes the evaluation model of driving fatigue based on Hidden Markov theory.The specific contents are as follows:1.Improving the process of training the hidden Markov model based on PSO algorithm.The traditional Baum-Welch algorithm is easy to make the model fall into local optimum has poor accuracy.So this paper proposes that introducing reasonably particle swarm algorithm into the process of training model and solves effectively trained result's dependence on initial model values.2.Study on HMM identification model with two states and single characterization index.The driver fatigue state is divided into levels: awake and fatigue.Choosing fatigue characteristic parameters PERCLOS as observation variables and establishing HMM identification model.Studying model theory using the Baum-Welch algorithm and improved algorithm respectively based on above conditions.3.Study on HMM identification model with three states and two characterization indexes.Considering the limitations of above model,expanding the driver fatigue state into three levels:awake,fatigue and serious fatigue.Choosing PERCLOS and PERLVO observation variables and using fuzzy clustering method to establish relevant model.4.Experimental verification and comparison analysis of results.Based on above theoretical research results,verifying the relevant theoretical research methods using the experimental data of driving simulator and carrying on the pertinence contrast analysis for improved training method combining experimental results.The results show that the improved HMM training method can solve efficiently the problem of choosing initial model values and the improved HMM model has better accuracy and stability.