SportRGDET: Scene prior and relation graph enhanced detection transformer for sports scenes

Object detection in sports videos holds significant application value in tactical analysis, athlete tracking, and intelligent broadcasting. However, existing methods face two key challenges when processing sports scenes: on one hand, target distribution in sports footage is strictly constrained by field boundaries, yet generic detectors uniformly allocate detection resources across the entire image, causing frequent false detections in off-field regions due to the visual similarity between spectators and athletes; on the other hand, objects such as athletes, balls, and referees in sports scenes form stable spatial co-occurrence patterns governed by game rules, yet existing methods predict each object independently, failing to leverage cross-category relational cues to assist the detection of difficult targets. To address these issues, this paper proposes SportRGDET (Sport Relation-Graph DEtection Transformer), a framework that achieves accurate detection through two core modules: the Scene Prior Guided Module (SPGM) jointly estimates the spatial probability distribution and boundary gradient map of the field region from multi-scale encoder features, fuses them to generate a hierarchical field confidence mask, and performs spatial modulation on the deformable self-attention weights of the encoder, which reallocates attention resources toward the effective field area and progressively suppresses off-field interference across layers, thereby reducing false detections while incurring only moderate additional computation; the Relation Reasoning Enhancement Module (RREM) treats each candidate object as a node during the detection decoding stage, adaptively learns inter-node association strengths through a relation encoder based on category embeddings and relative spatial positions of node pairs, constructs a dynamic relation graph, and bidirectionally propagates contextual information across nodes via graph attention mechanisms, enabling difficult targets to refine their predictions by leveraging positional and categorical cues from surrounding identified objects, enhancing detection robustness under occlusion and blur conditions. Experimental results demonstrate that SportRGDET achieves significant performance improvements on multiple sports detection benchmark datasets, validating the effectiveness of scene prior and relation reasoning mechanisms in sports scene object detection.