AI Driven Sensor Optimization For Marksman Posture Analysis

Sammanfattning

This thesis investigates whether full-body shooting kinematics can be used to predict burst-fireperformance and how those predictions could guide sensor optimization for marksman postureanalysis. Motion capture data from an Xsens Link motion capture suit and performance datafrom Saab’s Ground Combat Indoor Trainer were collected from 20 participants during a standardized standing burst-fire exercise, yielding 1 173 valid three-shot burst groups. Four classification targets were defined: experience level, objective performance level, burst-level outcome,and recoil direction. A Temporal Convolutional Network (TCN) was compared with a RandomForest (RF) baseline under leave-one-shooter-out cross-validation. SHAP was used to analyzewhich body segments and kinematic streams contributed most to these predictions.The results show that shooting kinematics contain strong information about stable shooter characteristics and more limited information about within-shooter burst quality. The best modelsachieved a macro F1-score of 0.912 for performance level, 0.756 for experience level, and 0.413for burst-level outcome. Recoil direction was moderately predictable, primarily at the shooterlevel, with a best binary macro F1-score of 0.692. Across most tasks, the TCN outperformedthe RF, especially when the predictive signal depended on temporal dynamics rather than largelystatic posture. SHAP analysis showed that the arm region dominated model attribution, that fulljoint-angle coverage outperformed a reduced ergonomic joint subset, and that angular acceleration was particularly informative for recoil-recovery prediction.These findings show that full-body motion capture can support predictive modeling of burst-fireperformance and that temporal models are well suited for capturing recoil-related movementpatterns. They also suggest that sensor optimization should preserve broad upper-body jointcoverage rather than focusing only on the primary shooting arm.