BVGA-Net: Lightweight Network Coupling BiLoop Vector-Mixed Gradient Aggregation for Quayside Crane Driver State Detection Against Complex Cockpit Interference

Driver state detection in port quayside cranes faces challenges from complex cockpit backgrounds and lightweight deployment needs. We proposed a lightweight detection network BVGA-Net based on an improved receptive-field attention mechanism and multi-level feature gradient fusion strategy, which was validated on a proposed benchmark dataset using cross-scenario division for generalization. To address back-ground interference in the cockpit, we designed an enhanced feature extraction encoder, where the vector-mixed receptive-field adaptive attention uses unfold operations and spatial attention to address parameter-sharing limitations in traditional convolutions. The parameter-free SimAM is integrated to suppress noise based on local self-similarity of images, improving feature representation capability without additional parameters. To balance multi-scale detection accuracy and computational efficiency, we proposed the bidirectional cross-domain loop aggregation (BCLA) strategy with receptive-field spatial features and introduced multi-level feature gradient fusion (MFGF) module. The BCLA method fuses multi-scale features through bidirectional loop paths to enhance multi-scale objects detection, while the MFGF module combines a global branch with a local branch via gradient fusion to reduce computational complexity and improve feature diversity in complex backgrounds. Experiments demonstrate that the method achieves an mAP50 of 87.7% with 3.76M parameters and balances accuracy and edge-deployment efficiency for safety-critical port operations.