Enhancing identity stability in sperm motility analysis via density-adaptive tracklet stitching with YOLOv12n and BoostTrack++

Accurate and reliable analysis of sperm motility is critical for assessing male fertility and improving treatment outcomes in assisted reproductive technologies. In this study, we propose an automated sperm motility analysis framework that integrates the YOLOv12n object detector, the BoostTrack++ multi-object tracker, and a novel Density-Adaptive Tracklet Stitching (DATS) post-tracking refinement module. The proposed framework is evaluated on the Hi-LabSpermTracking dataset under three complementary experimental scenarios designed to assess frame-level detection performance, temporal tracking consistency, and cross-subject generalization. Scenario I comparatively evaluates six contemporary detection models to identify the most suitable detector for microscopic sperm images. Scenario II evaluates multiple tracking algorithms under temporally separated training and testing conditions, while Scenario III assesses the final framework using videos acquired from previously unseen individuals. Experimental results demonstrate that the proposed framework achieves consistently strong detection and tracking performance across multiple evaluation metrics, including HOTA, MOTA, and IDF1. The proposed DATS module further reduces residual identity fragmentation by adaptively reconnecting fragmented tracklets according to local sperm density, improving trajectory continuity without modifying the underlying tracking architecture. Comprehensive comparisons with CNN-based, transformer-based, and end-to-end tracking approaches, together with statistical analyses, ablation experiments, and qualitative evaluations, demonstrate the robustness of the proposed framework under challenging microscopic sperm tracking conditions. Overall, the proposed framework provides an effective and practical solution for automated trajectory-based sperm motility analysis using conventional bright-field microscopy videos.