Multi-scale guided semantic enhancement for salient object detection in traffic scenes

Accurate detection of salient objects is a fundamental prerequisite for autonomous driving systems to support effective obstacle-avoidance decisions. However, directly applying generic saliency models to traffic scenarios suffers from a severe semantic gap, where visually prominent background distractors may obscure safety-critical targets. To address this issue, this paper proposes GSENet, a traffic salient object detection network based on global feature interaction and multi-scale semantic enhancement. Using a progressive encoder-decoder architecture, GSENet employs a hierarchical visual Transformer to capture long-range dependencies, alleviating the limited receptive fields of traditional convolutional neural networks. Meanwhile, a multi-scale semantic enhancement module enables perception of objects ranging from large foreground obstacles to small distant targets through multi-path parallel convolutional expansion. Furthermore, a cross-scale semantic interaction fusion module aligns global semantic flows via linear cross-attention and suppresses low-level feature noise through structure-aware gating. Deep supervision with a composite loss function further improves structural consistency across different resolutions. Quantitative experiments on TSOD10K show that GSENet achieves competitive performance, with an MAE of 0.0089 and a mean F-measure of 0.8582. Compared with VMamba, GSENet uses about 65% fewer parameters, yielding a favorable accuracy-efficiency trade-off.