EVALUASI KINERJA ALGORITMA MACHINE LEARNING DALAM KLASIFIKASI POSTUR DUDUK MENGGUNAKAN YOLOV8N-POSE

PERFORMANCE EVALUATION OF MACHINE LEARNING ALGORITHMS FOR SITTING POSTURE CLASSIFICATION BASED ON YOLOV8N-POSE KEYPOINTS

Abstract

Penggunaan perangkat digital yang meningkat memperpanjang durasi duduk dan berisiko menimbulkan gangguan muskuloskeletal akibat postur yang tidak ergonomis. Penelitian ini mengembangkan sistem klasifikasi postur duduk berbasis citra menggunakan YOLOv8n-Pose (Nano) untuk ekstraksi 17 titik sendi tubuh, yang kemudian diratakan (flattening) menjadi vektor 51 fitur numerik dan dinormalisasi dengan Standard Scaler. Dataset gabungan dari Roboflow dan Kaggle, yang terdiri atas 3.558 citra dengan rincian 1.044 good_form dan 2.514 bad_form, diuji pada lima algoritma yaitu Random Forest, KNN, SVM, MLP, dan Deep MLP, menggunakan skema hold-out 80:20 berbasis stratified sampling. Pada data uji, Random Forest mencatat performa generalisasi paling seimbang dengan akurasi 0,9775 dan Macro F1-Score 0,9728, sedikit mengungguli MLP Neural Network yang menunjukkan performa yang juga kuat, meskipun Random Forest sebelumnya menunjukkan indikasi overfitting pada fase pelatihan. Sistem kemudian diimplementasikan sebagai aplikasi web bernama SitSmart di Hugging Face Spaces. Penelitian ini mengindikasikan bahwa kombinasi ekstraksi koordinat pose dan model klasifikasi ensemble mampu memberikan hasil kompetitif untuk pemantauan ergonomi otomatis, meski validasi lapangan real-time masih diperlukan pada tahap lanjut.

Rising use of digital devices has extended sitting durations, thereby increasing the risk of musculoskeletal disorders from poor posture. This study develops an image-based sitting posture classification system using YOLOv8n-Pose (Nano) to extract 17 body keypoints, which are then flattened into a 51-feature numeric vector and normalized with Standard Scaler. A combined Roboflow and Kaggle dataset, consisting of 3,558 images with 1,044 good_form and 2,514 bad_form samples, was evaluated across five algorithms, namely Random Forest, KNN, SVM, MLP, and Deep MLP, using an 80:20 stratified hold-out split. On the test set, Random Forest achieved the most balanced generalization, with an accuracy of 0.9775 and a Macro F1-Score of 0.9728, narrowly ahead of MLP Neural Network, which also showed strong performance, although Random Forest had previously shown signs of overfitting during training. The best-performing model was then deployed as a web application called SitSmart on Hugging Face Spaces. These findings suggest that combining pose-based coordinate extraction with an ensemble classifier can yield competitive results for automated ergonomic monitoring, although real-world and real-time field validation remains necessary as a direction for future work.