Low-Visibility Vehicle-Road Environment Perception Based on the Multi-Modal Visual Features Fusion of Polarization and Infrared

As a key component of the automatic driving assistant system, vehicle-road environment perception technology is the basis of decision-making and the guarantee of safe driving. However, the existing vehicle-road environment perception technology based on vision mainly focus on the conditions of good illumination and weather and rely on visible light imaging equipment, which is not highly adaptable to special environments. In this paper, the key technologies of low-visibility vehicle-road environment perception are deeply studied by combining multi-mode visual imaging and deep learning, and a model of Multi-Branch Input Encoding-Decoding Network (MBINEDN) is proposed. In this model, the polarized degree image and infrared image of the same road scene are input simultaneously in the coding stage, and two branch networks are used to extract the multi-mode features. The lightweight network MobileNetV2 is used as the feature extractor to reduce the amount of model parameters, so as to retain the polarization and infrared feature of the image to the maximum extent. Finally, the two are combined and output in the decoding stage. Experiments show that using MBINEDN model to process two images can effectively improve the segmentation effect. At the same time, it shows that the effect of vehicle-road environment semantic segmentation based on multi-modal visual features is better than that of single modal features. Therefore, the application of multi-mode visual images to the segmentation of low-visibility vehicle-road scenes can provide a strong guarantee for the visual perception technology of vehicle safety driving assistance system.