Workshop on Multimodal Fusion enabled Embodied Intelligence of Autonomous Driving

Highlight: A cognitive learning framework enabling autonomous vehicles (AVs) to achieve human-like adaptability and safety in dynamic open-road environments is designed. By integrating (1) Spatial-Temporal Attention (STA) mechanisms to infer road-user intentions by dynamically prioritizing critical spatial regions and temporal segments; (2) Social Compliance Estimation via formalized "absolute right-of-way" (A_ROW) metrics and ROW-violation indices, ensuring socially acceptable interactions; (3) Deep Evolutionary Reinforcement Learning (DERL) combining Twin Delayed DDPG (TD3) with genetic algorithms to expand policy search space, avoid local optima, and balance safety, efficiency, and comfort. The framework demonstrates robust adaptation across traffic densities, bridging machine precision with human-like cognitive flexibility for safer, interpretable AV integration into real-world traffic.

Highlight: In simulation-based transportation systems—such as autonomous driving in adverse weather—decision optimization under noisy environments remains a core challenge. This talk introduces an Identification-Aware Bayesian Optimization framework, which goes beyond merely finding high-performing solutions to robustly identifying them. By balancing exploration and exploitation with adaptive acquisition strategies, the proposed method enhances decision reliability and sample efficiency. Applications in traffic signal control and trajectory planning demonstrate how this approach supports embodied intelligence in uncertain, multimodal scenarios.

Highlight: In real-world autonomous driving, vehicles must operate safely across a wide range of challenging visual conditions — from night-time urban roads to dim tunnels and rainy or foggy environments. Among these, low-light conditions are especially problematic, as they can severely compromise the performance of perception modules by obscuring critical visual cues, increasing noise, and reducing image contrast. Enhancing image quality in such scenarios is therefore essential for improving safety and reliability in autonomous systems. This presentation introduces ECAFormer, a lightweight and effective framework specifically designed for low-light image enhancement. By employing a novel dual cross-attention mechanism, ECAFormer enables mutual enhancement of semantic and visual features across multiple scales, effectively balancing global brightness correction and local detail preservation. Extensive experiments on benchmark datasets and real-world dark road scenes show that the model significantly improves visibility and robustness while maintaining computational efficiency.

Highlight: In modern urban traffic systems, complex traffic scenarios pose significant challenges to traffic management and intelligent transportation systems. To effectively address these challenges, it is essential to leverage multi-view and multimodal information for collaborative perception. This presentation explores how information can be acquired from various perspectives by combining different sensors and data sources, such as cameras, LiDAR, and in-vehicle sensors. It also focuses on how diverse modalities, such as images and videos, can be jointly analyzed and interpreted. This presentation introduces attention-based methods for the fusion and processing of multi-modal information, enabling systems to integrate data more effectively. Furthermore, the presentation will discuss how multi-modal data can be transformed into higher-level semantic understanding of traffic scenes, including modeling of pedestrians, vehicles, and road conditions. Finally, it will examine key challenges and future directions for collaborative perception in intelligent transportation.

Highlight: According to the road traffic injuries report of WHO, nearly 1.3 million people die in traffic accidents every year, and a high proportion of death results from driver attention distraction. Therefore, predicting where human driver should look is extremely important for advanced assisted driving systems and applications. This presentation discusses how to accurately predict human driver attention by fusing multiple data sources, such as cameras and in-vehicle sensors. This presentation introduces a method that predict human driver attention by simulating human driving experience accumulation procedure. Furthermore, this presentation also introduces another method that integrates multi-fold top-down guidance with the bottom-up feature.