Sanda Competition Detection Technology Integrating Deep Learning Image Segmentation and Object Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

To accurately evaluate the technical mastery and practical ability of Sanda athletes, the study innovatively proposes a Sanda competition detection method that combines deep learning image segmentation and target detection. This method first uses an improved Pyramid Scene Parsing Network (PSP-Net) to perform semantic segmentation of game video frames. Then, the Faster Region-based Convolutional Neural Network (Faster RCNN) and the Visual Geometry Group-19 (VGG19) network architecture are combined to detect hitting behaviors in key action frames, accurately detecting player actions and hitting behaviors in Sanda competitions. The proposed method exhibited superior stability, accuracy and generalization capabilities in different complex environments. In standard posture, fast action, noise and cross-subject generalization environments, the mean pose reconstruction errors after 150 rounds of training were only 2.97, 3.24, 4.52, and 3.81, respectively. In addition, the median dispersion scores for the four types of actions including front kick, side kick, rotating back kick and backhand boxing were 0.043, 0.047, 0.050 and 0.043, respectively. The fluctuation range was small, showing excellent recognition consistency for different action types. The method can achieve high-precision detection and stable evaluation under multiple environments, multiple action types and speed changes, thereby providing reliable technical support for intelligent referee assistance systems and sports training analysis.