Quantitative Evaluation of OpenCV Visual Tracking Algorithms for Drawn Markers Under Dynamic Facial Deformation

This study evaluates the eleven visual tracking algorithms available in OpenCV for tracking hand-drawn markers on deforming human faces. Using 41,641 multi-camera frames of seven subjects performing exaggerated phonetic and emotional movements, we assess each tracker’s deviation from an automated Blob Detection reference (four-marker sequences), its long-term marker identity retention (forty-marker sequences), and its throughput. Of the tested classical trackers, the Channel and Spatial Reliability Tracker (CSRT) performed best, retaining 96.2% of markers with a mean deviation of 1.94±1.29 pixels (approximately 0.46±0.31 mm) at 56.9 frames per second (FPS). The Kernelized Correlation Filter (KCF) matched this deviation at 445.4 FPS but failed under transient self-occlusion, retaining only 64.1%. Among the neural trackers, the Vision Transformer-based tracker (ViT) recorded the smallest deviation (1.06±1.03 pixels, approximately 0.25±0.25 mm) at 164.1 FPS, and the Distractor-Aware Siamese Region Proposal Network (DaSiamRPN) had the highest retention (92.4%, 1.35±1.26 pixels) at 11.6 FPS. Differences among the leading trackers were not significant at the subject level. The Minimum Output Sum of Squared Error (MOSSE) tracker reached about 12,500 FPS, but retained only 49.6%. Per-frame deviation and long-term retention are decoupled, meaning that both must inform tracker selection; scenario-based recommendations are provided.