Integrating Driving Behavior and Traffic Context Through Signal Symbolization for Data Reduction and Risky Lane Change Detection

A novel method for integrating driving behavior and traffic context through signal symbolization is presented in this paper. This symbolization framework is proposed as a data reduction method for naturalistic driving studies. Continuous sensor signals have been converted and reduced into sequences of symbols (chunks) using a sticky hierarchical Dirichlet process hidden Markov model and a nested Pitman-Yor language model. Then, co-occurrence chunking (COOC), the proposed integration method, has been applied to the driver behavior and the traffic context chunks. After the integration, COOC chunks have been associated with prototype driving scenes by using latent Dirichlet allocation. Finally, the translated sequence of chunks has been clustered into groups. Risky lane change detection experiments have been conducted with the symbolized data for evaluation purposes. A dataset comprised of 988 lane change scenes has been utilized for this process. Co-occurrence chunking with clustering provided the best risky lane change detection.