Human-Like Maneuver Decision Using LSTM-CRF Model for On-Road Self-Driving

In the near future, self-driving vehicles will be frequently tested in urban traffic, and will definitely coexist with human-driving vehicles. To harmoniously share traffic resources, self-driving vehicles need to respect behavioral customs of human drivers. Taking on-road driving for example, self-driving vehicles are supposed to behave in a human-like way to decide when to keep the lane and when to change the lane. This point, however, has not been well addressed in current on-road maneuver decision methods. In this paper, a human-like maneuver decision method based on Long Short Term Memory (LST-M) neural network and Conditional Random Field (CRF) model is proposed for on-road self-driving. Different from previous works, this paper considers the maneuver decision problem as a sequence labeling problem. Its input is a time-series vector which describes a period of neighboring traffic history, and its output is a one-hot vector indicates the suitable maneuver. The proposed model is trained on the NGSIM public dataset, which contains millions of driving maneuvers collected from thousands of human drivers. Simulations with manipulated conditions reveal human-like reasoning for maneuver decision inside the proposed model. Comparative experiments further demonstrate a better human-like performance achieved by the proposed method than that of previous methods.