Generating Natural Language Discription For Vehicle Trajectories Based On HMM

Automatically generating natural language description for vehicle trajectories is a research involves the conversion of visual information to the language information. It can be applied to describe traffic condition, and provide efficient transport services; it also helps psychologists to study human multi-modal conversion.This paper extracts two kinds of features which are the shape feature and the speed feature from vehicle trajectory, and extract basic words and word categories from manual annotated description based on hierarchical clustering, then use the 2-Gram language model to calculate the framework of vehicle trajectory description sentence, furthermore use the HMM to align the trajectory features and words, finally combine the words and description framework to generate description sentence.Trajectory classification is a key step of generating natural language description for vehicle trajectory. The HMM is better than other classifiers in processing of time series data, therefore it was selected for the classification of traffic trajectory. When HMM is used in the vehicle trajectory classification, the classification performance is affected by feature extraction, category design, and the HMM model parameters. Previous methods are not well suited for this particular task. This paper improves in three aspects:Firstly, Previous studies use original location or direction of the velocity to be the feature of the shape of a trajectory, however, these two features have their inherent defects. The original location builds a huge feature space, while the direction of the velocity ignores some important location information. This paper proposes a fitness feature extraction method, and it brings a significantly improvement to the performance of classification. Secondly, when designing the categories of the shape of trajectories, previous studies ignored the direction of a trajectory. This paper divides a shape to two categories based on different direction. It can obviously improve the classification performance. Thirdly this paper analyzes the effects of the choice of HMM topology and the number of hidden states on the classification results. Furthermore this paper analyzes the mapping between the abstract meaning of the hidden states and the physical properties of the vehicle trajectories. With this analysis, it is easy to choose the best HMM parameters for the vehicle trajectory classification task.