SCE-YOLO: Campus Abnormal Behavior Detection via Spatial-Frequency Enhancement and Dynamic Position Awareness

Detection of abnormal campus behavior is challenging because abnormal targets often exhibit large scale variations, small spatial extensions, frequent occlusions, complex background interference, and strong dependencies on relative spatial relationships. To address these issues, this study proposes SCE-YOLO, a campus-based abnormal behavior detection framework built on YOLO11n. The framework is designed from three complementary computational perspectives: spatial-frequency representation enhancement, dynamic spatial relationship modeling, and efficient multi-scale feature reconstruction. First, a C3k2-SFEM module is introduced to jointly model local texture information, directional edge responses, and global frequency components, thereby improving the representation of fine-grained abnormal cues, low-contrast targets, and irregular structures. Second, a C2DPB module incorporates Dynamic Position Bias into the attention mechanism, enabling feature interactions to be jointly determined by semantic similarity and relative spatial geometry. This design strengthens the modeling of human poses, target interactions, and long-range contextual dependencies under crowded and occluded conditions. Third, an Efficient Up-Convolution Block (EUCB) is embedded in the neck to reconstruct high-resolution spatial information through lightweight upsampling, depthwise convolution, and channel interaction, reducing boundary blurring and detail degradation during multi-scale feature propagation. In addition, the Campus Abnormal Behavior Detection Dataset (CABDD) constructed contains five representative abnormal-event categories, covering diverse viewpoints, target scales, illumination conditions, and occlusion levels. Experimental results show that SCE-YOLO achieves a Precision of 85.6%, a Recall of 71.7%, an mAP50 of 77.0%, and an mAP50:95 of 55.6% on CABDD, representing improvements of 4.5, 4.7, 4.6, and 4.1 percentage points, respectively, over the YOLO11n baseline. These results demonstrate that jointly enhancing feature representation, spatial interaction modeling, and multi-scale reconstruction can effectively improve abnormal-target recognition and localization in complex campus surveillance scenes.