Vehicle visual perception is essential for safe autonomous driving, especially in challenging low-visibility conditions. Polarimetric imaging has shown enhanced perception by improving target-background contrast and reducing glare. However, current research in polarimetric imaging for autonomous driving largely focuses on singular use of polarimetric features. Exploiting polarimetric information to enhance vehicle visual perception in low-visibility scenarios remains a critical challenge. This paper addresses this challenge by integrating multiple polarimetric parameter characterization with a deep learning model for semantic segmentation visual perception task, which is called TransWNet. The model combines Convolutional Neural Networks (CNN) and Transformer architectures, extracting features and contextual information of both Degree of Polarization (DoP) and Angle of Polarization (AoP) comprehensively during the encoding phase. In the decoding phase, it incorporates skip connections to effectively retain multimodal polarimetric information across deep and shallow features, generating output through feature fusion. To the best of our knowledge, this is the first work that jointly exploits DoP and AoP for vehicle scene understanding in degraded visibility. Experimental results demonstrate that TransWNet, by effectively leveraging multimodal polarimetric information, achieves significantly better performance in semantic segmentation of low-visibility traffic scenes, with marked improvements in mIoU, mPA, and Accuracy over all single-feature baselines. Compared with the baseline method, TransWNet improves the Accuracy by 3.44%.

Vehicle platooning is one of the advanced driving applications expected to be supported by the 5G vehicle to everything (V2X) communications. It holds great potentials on improving road efficiency, driving safety and fuel efficiency. Apart from the organization and internal communication of the platoons, real-time prediction of surrounding road users (such as vehicles and cyclists) is another critical issue. While artificial intelligence (AI) is receiving increasing interests on its application to trajectory prediction, there is a potential problem that the pre-trained neural network models may not well fit the current driving environment and needs online fine-tuning to maintain an acceptable high prediction accuracy. In this paper, we propose a digital twin based real-time trajectory prediction scheme for platoons of connected intelligent vehicles. In this scheme the head vehicle of a platoon senses the surrounding vehicles. A LSTM neural network is applied for real-time trajectory prediction with the sensing outcomes. The head vehicle controls the offloading of the trajectory data and maintains a digital twin to optimize the update of LSTM model. In the digital twin a Deep-Q Learning (DQN) algorithm is utilized for adaptive fine tuning of the LSTM model, to ensure the prediction accuracy and minimize the consumption of communication and computing resources. A real-world dataset is developed from the KITTI datasets for simulations. The simulation results show that the proposed trajectory prediction scheme can maintain a prediction accuracy for safe platooning and reduce the delay of updating the neural networks by up to 40%.

Advanced sensor technology and emerging Internet of Vehicles (IoV) have significantly accelerated the realization of autonomous driving. In operating an autonomous vehicle, traffic information needs to be collected and processed under strict delay constraints. But an individual vehicle equipped with few types of sensors and limited computation resources could not achieve precise environmental awareness and real-time information processing. Vehicular cooperative sensing and edge computing are promising approaches to address these problems. However, various types of sensing tasks and heterogeneous smart vehicles with different computing power make the cooperation between vehicles a complicated problem. To cope this problem, we form multiple vehicles into platoons, and design a novel cooperative sensing architecture. Moreover, we fully exploit unoccupied computation resources of smart vehicles, and propose a vehicular edge serving scheme, which jointly schedules cooperative sensing and task offloading. Numerical results demonstrate that our proposed scheme outperforms traditional approaches with lower delay costs.