Research & Papers

DefocusTrackerAI: YOLOv9-based framework tracks particles with sub-0.4 pixel accuracy

New deep learning model detects defocused particles with 0.1–0.4 pixel precision, even in dense sprays.

Deep Dive

Researchers introduced DefocusTrackerAI, a generalized deep-learning framework for automatic detection and position estimation of defocused particle images. Built on YOLOv9, it outperforms Faster R-CNN and state-of-the-art algorithms, achieving higher recall and uncertainty between 0.1 and 0.4 pixels at particle densities up to N_s=0.5. The model handles astigmatic and non-astigmatic particles under varied lighting, and works on fluorescence and shadowgraph data. A pre-trained version is available for three-dimensional defocusing particle tracking.

Key Points
  • YOLOv9 model achieves 0.1–0.4 pixel uncertainty for particle densities up to N_s=0.5, outperforming Faster R-CNN and existing algorithms.
  • Framework works across multiple optical setups (astigmatic, non-astigmatic) and lighting conditions, validated on fluorescence and shadowgraph data.
  • Pre-trained DefocusTrackerAI model is publicly available as a ready-to-use tool for high-accuracy automated particle detection.

Why It Matters

Automates high-precision particle tracking in fluid dynamics and biomedical imaging, enabling faster, accurate 3D analysis at scale.