Abstract:

In existing end-to-end deep learning-based autonomous driving frameworks， there is a common problem of low accuracy in planning and control prediction， often due to the single-source input data and inability to balance spatial and temporal information. To better reflect the impact of the historical interaction process between the ego vehicle， environment， and traffic participants on the current decision-making in virtual simulation testing， this paper designs a multi-level spatiotemporal attention long short-term memory network for vehicle motion planning in autonomous driving simulation environment. The algorithm extracts and represents deep abstract information of the autonomous driving environment and realizes end-to-end vehicle motion control in the simulation platform. Firstly， a convolutional module is used to extract spatial features of a single image at a specific moment using the historical continuous video frame sequence of RGB simulation data acquired by the forward-facing camera model as input. Secondly， the LSTM module is used to fuse the spatial information of the image across historical moment to obtain temporal contextual features. Additionally， to enhance the ability to extract spatiotemporal key information and accelerate network convergence， a spatiotemporal attention mechanism is applied in the fusion part of the multi-level spatiotemporal features. The proposed method is tested and validated on the Carla simulation platform. The experimental results show that the proposed method can more accurately simulate human driving decision-making behavior compared to the single spatiotemporal algorithm.

Key words: vehicle motion planning, end-to-end, space-time attention, deep learning, simulation, LSTM