Automatic Mining of Vehicle Behaviors with an Unknown Number of Categories
Ying Liu, Hao Zhang, Huadong Meng, Xiqin Wang · 2008
Automatic mining of vehicle behaviors from raw data collected by multiple sensors provides meaningful qualitative descriptions of the vehicle status. These qualitative behavior descriptions can be used in scenario parsing and have further applications in vehicle surveillance and frontal collision warning systems. In current approaches, the number of behavior categories is supposed to be known, or need to be manually explored every time the training data is changed. In this paper, the authors use Hidden Markov Model to symbolize the vehicle behaviors and adopt the cross-validated likelihood with penalty for complexity to select the number of hidden states. Appropriate number of behavior categories is selected automatically, and those behaviors are decided at the same time. Real data experiments demonstrate the effectiveness of this approach.