Traffic scene prediction for autonomous driving using CNN and LSTM networks considering consecutively interacting vehicles

Abstract

Recently, autonomous driving technology on highway situation, semi-autonomous driving, is already commercialized by a number of automotive company. The core technology of the autonomous driving vehicles that leads the fully autonomous driving generation, however, is to extend the technology on the highway situation to urban traffic scenario. Google and Tesla cannot prevent accident of the autonomous vehicles in urban traffic scenario due to inaccurate future prediction technology also. In complex traffic situation, there are a number of actively interacting agents compare to highway situation. Conventional future prediction technology for autonomous vehicles are concentrated on the mono-vehicular prediction so that the accuracy in multi-vehicle environment is quite low because of existence of a number of traffic participants interacting each other. Therefore, this thesis proposes the deep learning architecture for multi-vehicle prediction by introducing the concept of multi-channel occupancy grid map which can represent the overall surrounding traffic scene around the ego vehicle. By doing so, all of the surrounding vehicles are predicted as an integrated manner so that interaction can be predicted. The predicted future OGMs up to 4 seconds are generated via convolutional neural network and long short-term memory networks using historical multi-channel OGM input with static layer channel and dynamic layer channel. While previously developed future prediction algorithm was limited to the one target vehicle, proposed deep learning architecture can provide effective prediction on multiple surrounding vehicles simultaneously by extracting interaction feature from multi-channel OGM. The accuracy of the proposed deep learning architecture trained with 36,000 training set was around 94.69% in 2 seconds of prediction horizon. Also, the proposed algorithm is validated with several case studies on various scenario. Based on these results, the proposed algorithm is expected to be the leading technology to extend the autonomous vehicle technology on high-way like environment into urban traffic situation.