A Two-Stream Asymmetric Data Fusion Approach with Low Data Redundancy for Object Detection in Autonomous Driving
Yuchuan Fu, Xiaojian Niu, Yu‐Ting Huang, Changle Li · Advances in transdisciplinary engineering · 2025
Accurate environmental perception is crucial for autonomous driving, empowering self-driving vehicles to make intelligent decisions and effectively control their actions. While traditional methods often rely on multi-sensor data fusion to overcome the inherent limitations of individual sensors, many existing symmetric unidirectional fusion techniques fail to fully leverage the relationships between different data modalities, resulting in significant redundancy in the fused features. Additionally, these methods often struggle with robustness in complex scenarios, such as adverse weather conditions, leading to challenges when sensor data is lost. To tackle these issues, we propose a two-stream asymmetric data fusion approach (TSAFusion) that significantly enhances efficiency and reduces redundancy. Our method integrates multimodal features within a shared Bird’s Eye View (BEV) representation space and employs bidirectional fusion to thoroughly explore the interrelationships between different modalities. This not only minimizes redundant information but also improves overall fusion efficiency. Furthermore, we incorporate an attention mechanism that enhances the model’s robustness in varying weather conditions. Extensive simulation results underscore the advantages of our proposed algorithm, demonstrating superior detection accuracy, enhanced fusion performance, and faster convergence compared to existing methods.