Analisis Performa Model Deteksi Pose Dan Pelacakan Pegawai Pada Sistem Penghitung Durasi Kerja

Danial, Muhammad Danial Ervany (2026) Analisis Performa Model Deteksi Pose Dan Pelacakan Pegawai Pada Sistem Penghitung Durasi Kerja. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 2040221019-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

PT. Petrokimia Gresik sebagai perusahaan solusi agroindustri skala nasional membutuhkan data durasi kerja aktual Pegawai di kantor Gudang Multiguna, Departemen Pergudangan dan Pengantongan, namun belum memiliki sistem pencatatan yang mampu merekam aktivitas kerja secara objektif dan realtime. Infrastruktur kamera CCTV yang sudah terpasang selama ini hanya berfungsi sebagai pengawas pasif tanpa menghasilkan data produktivitas yang terukur, sehingga manajemen tidak memiliki acuan valid untuk mengevaluasi efektivitas jam kerja setiap shift secara berbasis data. Penelitian ini mengembangkan sistem penghitung durasi kerja Pegawai berbasis computer vision dengan memanfaatkan infrastruktur CCTV yang sudah ada, sekaligus melakukan analisis komparatif terhadap sebelas algoritma yang mencakup tiga komponen utama yaitu deteksi objek YOLOv8, YOLOv11, YOLOv12, pelacakan objek ByteTrack, DeepSORT, BoTSORT, StrongSORT, dan estimasi pose MediaPipe, YOLOPose, MoveNet, OpenPose untuk menemukan konfigurasi model dengan performa terbaik pada kondisi lingkungan kerja nyata. Variabel yang diukur meliputi akurasi, FPS, latensi inferensi, serta penggunaan CPU, GPU, RAM, dan VRAM, dengan pemilihan model terbaik menggunakan metode TOPSIS yang mempertimbangkan seluruh variabel performa secara simultan. Hasil pengujian menunjukkan bahwa YOLOv11 unggul dalam kecepatan dan efisiensi komputasi dengan latensi inferensi konsisten di bawah 43 ms dan penggunaan GPU terendah berkisar 39,47%-42,39%, sementara YOLOv8 menghasilkan confidence score tertinggi mendekati 90% dengan akurasi paling konsisten. Pada komponen pelacakan objek, DeepSORT terbukti paling andal dalam mempertahankan identitas target dengan jumlah ID switch paling minim sebanyak 7 objek unik dan lifetime pelacakan terpanjang mencapai 899,8 frame, sedangkan ByteTrack mengungguli seluruh model pada kecepatan dengan 48 FPS dan beban GPU terendah 27,3%. Pada komponen estimasi pose, MediaPipe mendominasi secara komprehensif dengan akurasi tertinggi 77,87%, mampu memetakan rata-rata 25,7 dari 33 keypoint valid sekaligus mempertahankan kecepatan realtime 47,2 FPS. Berdasarkan analisis TOPSIS, ByteTrack meraih skor preferensi tertinggi pada kategori pelacakan objek sebesar 0,8033, sedangkan MediaPipe dan YOLOv11 masing-masing terpilih sebagai model terbaik pada kategori estimasi pose dan deteksi objek.
=================================================================================================================================
PT. Petrokimia Gresik, a national-scale agro-industrial solutions company, requires data on employees’ actual working hours at the Multipurpose Warehouse office within the Warehousing and Bagging Department; however, it does not yet have a tracking system capable of recording work activities objectively and in real time. The existing CCTV camera infrastructure has only functioned as a passive surveillance system without generating measurable productivity data, leaving management without a valid basis for evaluating the effectiveness of each shift based on data. This study develops a computer vision-based employee work duration tracking system by leveraging the existing CCTV infrastructure, while also conducting a comparative analysis of eleven algorithms covering three main components: object detection YOLOv8, YOLOv11, YOLOv12, object tracking ByteTrack, DeepSORT, BoTSORT, StrongSORT, and Pose Estimation MediaPipe, YOLOPose, MoveNet, OpenPose to identify the model configuration with the best performance under real-world work conditions. The measured variables include accuracy, FPS, inference latency, as well as CPU, GPU, RAM, and VRAM usage, with the best model selected using the TOPSIS method, which considers all performance variables simultaneously. Test results show that YOLOv11 excels in speed and computational efficiency, with inference latency consistently below 43 ms and the lowest GPU utilization ranging from 39.47% to 42.39%, while YOLOv8 produces the highest confidence scores approaching 90% with the most consistent accuracy. In the object tracking component, DeepSORT proved to be the most reliable in maintaining target identity with the fewest ID switchesjust 7 unique objects and the longest tracking lifetime of 899.8 frames, while ByteTrack outperformed all models in speed at 48 FPS and the lowest GPU load of 27.3%. In the Pose Estimation component, MediaPipe dominated comprehensively with the highest accuracy of 77.87%, capable of mapping an average of 25.7 out of 33 valid keypoints while maintaining a real-time speed of 47.2 FPS. Based on the TOPSIS analysis, ByteTrack achieved the highest preference score in the object tracking category at 0.8033, while MediaPipe and YOLOv11 were each selected as the best models in the Pose Estimation and object detection categories, respectively.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Computer vision, YOLO ,Object Tracking , Pose Estimation ,Timer , TOPSIS. Computer vision, YOLO, Object Tracking, Pose Estimation, Timer, TOPSIS.
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing. T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5105.546 Computer algorithms T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK6592.A9 Automatic tracking. T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Vocational > 36304-Automation Electronic Engineering
Depositing User:	Muhammad Danial Ervany
Date Deposited:	14 Aug 2026 06:01
Last Modified:	14 Aug 2026 06:01
URI:	http://repository.its.ac.id/id/eprint/144343

Actions (login required)

View Item