FDGR-Net: a two-stage framework for unified assistive-driving multi-task perception

Advanced Driver Assistance Systems (ADAS) require holistic perception of driver state, vehicle behavior, and surrounding traffic context, yet existing studies usually model these tasks independently despite their strong real-world correlations. Multi-task learning provides a natural way to jointly optimize related ADAS perception tasks, but most current approaches mainly focus on semantically homogeneous task groups and are insufficient for unified heterogeneous assistive-driving perception. A key challenge in this setting is the conflict between task commonality and task specificity, since driver-centric tasks rely on fine-grained cues while environment-centric tasks depend more on global scene structure, making naive shared representations prone to feature entanglement and negative transfer. Another limitation of existing methods is that they usually exploit task relationships only implicitly through shared features or simple fusion, without explicitly modeling the structured dependencies that could support effective positive transfer. To address these issues, we propose FDGR-Net , a novel two-stage framework for unified assistive-driving multi-task perception, in which the first stage performs Task-aware Feature Decomposition to separate task-sensitive characteristics from shared representations and establish a stable discriminative backbone. Based on this foundation, the second stage introduces Cross-task Fusion via Task Graph Reasoning, which explicitly models inter-task relationships and enables adaptive information propagation to selectively exploit beneficial cross-task cues. Extensive experiments demonstrate that the proposed progressive design effectively alleviates representation conflict, enhances structured task collaboration, and consistently improves the overall performance of multiple heterogeneous ADAS-related perception tasks.