A Lightweight Extraction Method for Human Skeletal Joint Points of Wind Tunnel Pilots Based on GST

Wind tunnel skydiving training, as an essential means of skydiving and aviation training, relies on efficient posture capture to ensure training effectiveness. However, due to factors such as complex wind tunnel backgrounds,pilot body occlusion, small joint targets, and high model complexity, the real-time extraction accuracy of the current model needs to be improved. To this end, we propose a lightweight method based on Gst to extract the critical skeleton points of wind tunnel pilots. This method redesigns the lightweight keypoint extraction model YOLOv8nPose-Gst, an improved version of YOLOv8nPose aimed at improving wind tunnel pilot attitude detection accuracy and efficiency.By introducing the CoordAtt attention mechanism,the model is better at focusing on small keypoint targets, and combined with the HeadGst convolutional head design, it effectively reduces computational complexity while maintaining high detection accuracy. In addition, this article adopts the NmPIoU loss function to accelerate training and further improve model performance. Our method achieved significant results across multiple metrics on the wind tunnel pilot dataset. Compared with the baseline model YOLOv8nPose, map@50, and map@50-95, The values reached 0.801 and 0.564, 2.2 percentage points and 3.3 percentage points higher than the original model. Meanwhile, the model complexity GFLOPs is 7.8, a decrease of 1.4; The number of parameters has also decreased from 3.289 to 2.932.