DF-PLSTM-FCN: A Method for Unmanned Driving Based on Dual-Fusions and Parallel LSTM-FCN

摘要

Learning algorithms are increasingly being applied to behavioral decision systems for unmanned vehicles. In multi-source road environments, it is one of the key technologies to solve the decision-making problem of driverless vehicles. This paper proposes a parallel network, called DF-PLSTM-FCN, which is composed of LSTM-FCN-variant and LSTM-FCN. As an end-to-end model, it will jointly learn a mapping from the visual state and previous driving data of the vehicle to the specific behavior. Different from LSTM-FCN, LSTM-FCN-variant provides more discernible features for the current vehicle by introducing dual feature fusions. Furthermore, decision fusion is adopted to fuse the decisions made by LSTM-FCN-variant and LSTM-FCN. The parallel network structure with dual fusion on both features and decisions can take advantage of the two different networks to improve the prediction for the decision, without the significant increase in computation. Compared with other deep-learning-based models, our experiment presents competitive results on the large-scale driving dataset BDDV.