Robust autonomous driving control using deep hybrid-learning network under rainy/snown conditions

The study introduces a groundbreaking two-stage deep hybrid learning architecture, Robust Autonomous Driving Control (RADC), designed to address the formidable challenge of ensuring safe and efficient autonomous driving in adverse weather conditions, including heavy rain and snow, in complex scenarios. In the first stage, our proposal utilizes an encoder as a variational autoencoder (VAE) model. This encoder leverages the VAE to extract feature information from the perceptual data surrounding the environment. Moving on to the second stage, we build a decoder as an Inception-Bidirectional Long Short-Term Memory (IBL) model. This decoder combines the VAE latent features obtained in stage one with additional control vehicle information tasks, including steering, speed, etc. Our approach involves predicting driving behavior along a predetermined route, allowing autonomous vehicles to stay centered on the road, simulating diverse driving scenarios, and achieving significant reductions in route-following time. This framework utilizes deep hybrid learning methods and harnesses Nvidia GPU capabilities to evaluate the effectiveness of InceptionNet, ResNet-50, MobileNet, DenseNet, and VGG16 convolutional neural networks. It enhances vehicle route following with improved metrics such as reduced inference time, heightened accuracy, and minimized lane changes in training. The RADC demonstrates exceptional performance, notably achieving a mean square error of 0.0464, 0.0346, and a 6-millisecond inference time in challenging weather like hard rain and snow. The proposed model is validated through comprehensive experiments in the Airsim environment, showcasing notable interpretability, generalization, and robustness capabilities. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.