Sensor-free, Camera-only: Low-cost multi-angle CPR skill assessments method based on computer vision

Cardiopulmonary Resuscitation (CPR) is critical for survival following Cardiac Arrest (CA), particularly in out-of-hospital settings. However, current CPR skill assessments often rely on simulators with manual scoring, leading to inconsistencies and inefficiencies. While artificial intelligence (AI) has shown promise, many AI-based methods still depend on expensive, sensor-laden equipment. This study introduces a novel, sensor-free approach for CPR evaluation, aiming to provide a cost-effective solution for wider accessibility. We propose the LCSG-YOLO model, an enhanced version of the YOLO11 algorithm, coupled with the Kalman-Hungarian-Confidence-Temporal (KCHT) matching algorithm. The KCHT algorithm refines target tracking and assignment, improving detection precision and system efficiency by addressing challenges in dynamic object tracking and ambiguous assignments. By integrating these techniques, our system analyzes CPR training videos to assess compression position, frequency, and depth. We calculate standardized mean times for compressions and respirations, determine operation frequencies, and extract grayscale and jitter signals from compression regions using image processing and Fast Fourier Transform (FFT). This allows for accurate evaluation of compression depth and frequency characteristics. The LCSG-YOLO approach offers a practical, computer vision-based solution for CPR training and assessment, potentially expanding CPR education in resource-limited environments. The results show that the LCSG-YOLO model outperforms YOLO11s and YOLO11n, with improvements of 6.4% and 2.3% in chest compression detection box precision and 4.2% and 6.4% in artificial respiration detection recall, as detailed in our comparative performance analysis. The KCHT algorithm achieves detection accuracies of 98.6% and 98.7% for standard hand positioning during chest compressions and artificial respiration, respectively.