Projectile extraction from high-speed video under muzzle-blast smoke via YOLOv8 detection and GMM segmentation

The velocity of projectiles at specific distances from the muzzle plays a crucial role in ballistic research as well as forensic firearm identification. However, the considerable muzzle-blast smoke generated at the instant of firing often causes severe visual occlusion, posing significant challenges to accurate measurement. To address this issue, we introduce a projectile recognition and extraction pipeline based on high-speed video analysis. In the first stage, the YOLOv8 deep learning detector performs coarse localization by providing bounding boxes of candidate projectiles. Within each detection window, a Gaussian Mixture Model (GMM) is then fitted to pixel intensity features, and parameters are estimated by the Expectation–Maximization algorithm to obtain posterior probability distributions of pixel categories. This enables an effective separation of projectile, smoke, and background. Posterior thresholding and morphological refinement are subsequently applied to generate a clean binary mask, from which two-dimensional projectile centroids are extracted. Finally, stereo triangulation reconstructs the 3D trajectory, and inter-frame displacements yield instantaneous velocity estimates. Experimental results demonstrate that the proposed method reliably and efficiently identifies and extracts projectiles even under dense smoke interference. Compared with measurements obtained in clear conditions, the relative deviation of velocity remains below 1%, confirming the robustness and effectiveness of the approach.