SISTEM DETEKSI MANUSIA PADA AREA TERBATAS MENGGUNAKAN YOLOV7 BERBASIS CCTV

Abstrak

Pengawasan keamanan gedung kampus berbasis Closed-Circuit Television (CCTV) pada umumnya masih bergantung pada pemantauan manual sehingga rentan terhadap kelelahan operator dan keterlambatan respons. Penelitian ini bertujuan merancang dan mengimplementasikan sistem deteksi manusia otomatis pada area terbatas gedung kampus menggunakan You Only Look Once versi 7 (YOLOv7), serta mengevaluasi kinerjanya pada variasi pencahayaan dan jumlah objek. Zona pengawasan dibatasi dengan Region of Interest (ROI) berbentuk poligon yang dipadukan dengan parameter waktu operasional; pelanggaran ditetapkan apabila titik tengah bounding box manusia berada di dalam ROI di luar jam operasional. Sistem dibangun dengan Python, PyTorch, dan OpenCV menggunakan bobot YOLOv7 pralatih pada dataset COCO tanpa pelatihan ulang. Pengujian memakai 18.000 frame sampel dari 50 jam rekaman CCTV pada kondisi siang dan malam (inframerah). Hasil menunjukkan precision 99,47%, recall 98,61%, F1-score 99,04%, mAP@0,5 96,45%, dan kecepatan rata-rata 47 FPS pada GPU NVIDIA GeForce RTX 3050 Ti. Rerata confidence score menurun dari 87,4% (siang) menjadi 64,2% (malam), dan tingkat deteksi menurun seiring bertambahnya jumlah objek. Dari 150 kejadian pelanggaran, 142 (94,7%) teridentifikasi dengan waktu respons rerata 0,8 detik. Akurasi penentuan posisi terhadap ROI paling rendah di area perbatasan (94,40%). Hasil ini menunjukkan bahwa integrasi YOLOv7, ROI, dan parameter waktu layak digunakan sebagai alat bantu pemantauan otomatis pada lokasi uji.

Conventional Closed-Circuit Television (CCTV) surveillance in campus buildings still relies largely on manual monitoring, which is prone to operator fatigue and delayed response. This study designs and implements an automated human detection system for restricted campus areas using You Only Look Once version 7 (YOLOv7) and evaluates its performance under varying lighting conditions and numbers of people per frame. Surveillance zones are constrained by a polygonal Region of Interest (ROI) combined with an operating-hours parameter; a violation is registered when the center point of a person's bounding box lies inside the ROI outside operating hours. The system was built with Python, PyTorch, and OpenCV using pretrained YOLOv7 weights from the COCO dataset without retraining. Testing used 18,000 sampled frames from 50 hours of CCTV footage recorded in daytime and at night (infrared mode). The system achieved a precision of 99.47%, recall of 98.61%, F1-score of 99.04%, mAP@0.5 of 96.45%, and an average speed of 47 FPS on an NVIDIA GeForce RTX 3050 Ti GPU. The mean confidence score dropped from 87.4% (day) to 64.2% (night), and the detection rate declined as the number of people per frame increased. Of 150 violation events, 142 (94.7%) were identified with a mean response time of 0.8 s. Spatial classification accuracy was lowest near the ROI boundary (94.40%). These results indicate that combining YOLOv7, ROI, and time parameters is feasible as an automated monitoring aid at the single test site.