Adaptive multimodal fusion with semantic priority scheduling for cooperative 3D detection and communication methods

In response to the urgent demand for high-precision 3D object detection and efficient communication in intelligent transportation systems and autonomous driving, this paper proposes an Adaptive Multimodal Fusion and Semantic Priority Scheduling approach for collaborative 3D detection and communication (AMFS-C3D). During the fusion of multimodal sensor data (from cameras, LiDAR, and radar), a graph neural network is introduced to structurally encode the integrated 3D features, thoroughly capturing spatial interactions among targets; a variational autoencoder then compresses these high-dimensional graph embeddings, retaining essential semantic information even under bandwidth constraints and significant channel noise. To further reduce latency and increase communication efficiency, the system employs an adaptive priority scheduling mechanism that dynamically allocates bandwidth based on real-time network load and target importance, ensuring the timely transmission of critical objects. Experimental results show that AMFS-C3D significantly outperforms comparative methods in key metrics such as mean Average Precision (mAP), recall, semantic fidelity, and transmission latency. Under the same mAP threshold, AMFS-C3D reduces bandwidth requirements by over 10 % on average, and under high signal-to-noise ratio conditions, it improves detection accuracy and recall by 3–5 percentage points. Moreover, AMFS-C3D demonstrates superior applicability and robustness across diverse network loads and channel conditions.