A Framework for Driving Intention Estimation in Real-World Scenarios

Existing research on driving intention has focused on the estimation of single driving intention, such as whether a lane change will be made. In a real-world environment, safe driving requires a more comprehensive estimation for driving intention, such as rapid deceleration, changing lane, and turning. This study presents a complete framework for driving intention estimation in the real world, which consists of three parts: definition of the driving intention set, analysis of factors that influence driving intention estimation, and the input data set for driving intention estimation. The Berkeley open-source data set-BDD100K is used as the data source. Considering the typical time series characteristics of driving intention estimation, we use LSTM (long short-term memory) neural network as an advanced algorithm for modeling. In combination with the framework, LSTM is shown to be a promising model for estimating vehicle driving intentions in a real traffic environment.