Salient Object Detection of Dynamic Night Scenes via Bio-Inspired Spotlight Attention and Hierarchical Edge-Texture Fusion

The perception of night scenes is of crucial importance for driving safety. In the dimly lit night environment, as the visibility of objects decreases, both experienced and inexperienced drivers often struggle to fully notice the objects closely related to the driving task. Moreover, because the contours of many objects are blurred in dim night, locating and detecting objects are much more difficult than that in daytime scenes, especially for the small traffic objects, which undoubtedly greatly increases the potential road hazards. Till now, there are few studies specifically focusing on the night object detection based on driver’s attention. This research is dedicated to solving the detection problem of significant objects in night scenes, particularly small salient objects. First, we constructed a Night Eye-Tracking Object Detection Dataset (NETOD), which can provide a benchmark for research on attention-driven object detection in night scenes. Then, we proposed a salient object detection model for night traffic scenes, named NS-YOLO. NS-YOLO integrates a Bio-Inspired Spotlight Attention Module (BSAM) that combines bottom-up feature enhancement with top-down semantic guidance to accurately localize salient objects. Additionally, a hierarchical multi-scale detection architecture is introduced, leveraging cross-layer feature pyramid and dynamic upsampling to enhance the detection of small objects. The experimental results on the NETOD dataset show that the proposed salient small object detection model for night traffic scenes achieved mean Average Precision (mAP) value of 93.0%, outperforming other advanced models. It has important potential application values in driver assistance, danger warning, and other aspects, and is expected to significantly improve the safety and intelligence of night driving. Beyond technical advancements, this work highlights the necessity of human-centric attention mechanisms in autonomous systems, paving the way for safer and more interpretable AI-driven vehicles.