Pendekatan Deep Learning Untuk Deteksi Kejadian Anomali Pada Anak Taman Kanak-Kanak Menggunakan CCTV

Panjaitan, Giobert Boyhosea (2026) Pendekatan Deep Learning Untuk Deteksi Kejadian Anomali Pada Anak Taman Kanak-Kanak Menggunakan CCTV. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221310-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (13MB) | Request a copy

Abstract

Kejadian anomali seperti jatuh, memanjat, berkelahi, dan akrobatik merupakan risiko cedera serius pada anak usia dini (3–6 tahun) di lingkungan Taman Kanak-Kanak (TK) karena koordinasi motorik yang masih berkembang dan tingginya aktivitas fisik yang spontan. Pengawasan berbasis CCTV secara konvensional masih bergantung pada pemantauan manual oleh manusia, yang tidak efisien serta rentan terhadap kelelahan dan human error.
Penelitian ini mengembangkan sistem deteksi anomali otomatis berbasis deep learning untuk pemantauan keamanan anak TK menggunakan rekaman CCTV sudut atas (top-down). Dataset dikumpulkan secara mandiri dari rekaman CCTV TK Al-Hikmah Pasuruan dan terdiri dari 2.227 klip video yang diklasifikasikan ke dalam lima kelas: Normal, Jatuh, Memanjat, Berkelahi, dan Akrobatik. Sistem yang diusulkan mengadopsi arsitektur late fusion tiga model paralel: Model Pose menggunakan BiLSTM dengan Temporal Attention untuk memproses urutan keypoint tubuh yang diekstraksi oleh MediaPipe, Model Appearance Frozen menggunakan LSTM dengan Temporal Attention atas fitur ResNet-50 yang dibekukan, dan Model Appearance Fine-tuned menggunakan ResNet-50 layer4 yang dilatih ulang dikombinasikan dengan LSTM. Bobot late fusion ditentukan melalui grid search pada validation set.
Pada evaluasi multikelas, Fusion 3-Way mencapai F1-score makro sebesar 0,6441 dengan akurasi 70,6%. Pada evaluasi biner (Normal vs. Anomali), Fusion App+AppFT mencapai F1-score makro 0,8238 dengan akurasi 82,6%. Pengujian generalisasi pada dataset fall detection publik (GMDCSA + LE2I) menghasilkan F1-score 0,8904 dengan akurasi 94,7%. Seluruh konfigurasi fusion memenuhi target real-time dengan kecepatan inferensi ≥25 FPS.
===================================================================================================================================

Anomalous incidents such as falls, climbing, fighting, and acrobatics pose serious injury risks for early childhood (ages 3–6) in kindergarten environments due to still-developing motor coordination and inherently spontaneous physical activity. Conventional CCTV systems remain passive tools reliant on manual monitoring, which is inefficient and susceptible to human error and fatigue.
This research develops an automated anomaly detection system using deep learning for child safety monitoring in kindergartens via top-down CCTV footage. The dataset was independently collected from CCTV recordings at TK Al-Hikmah Pasuruan and consists of 2,227 video clips classified into five categories: Normal, Fall, Climbing, Fighting, and Acrobatics. The proposed system adopts a late fusion architecture of three parallel models: a Pose Model using BiLSTM with Temporal Attention over MediaPipe body keypoint sequences, an Appearance Frozen Model using LSTM with Temporal Attention over frozen ResNet-50 features, and an Appearance Fine-tuned Model using fine-tuned ResNet-50 layer4 combined with LSTM. Fusion weights are determined via grid search on the validation set.
In multiclass evaluation, the 3-Way Fusion achieved a macro F1-score of 0.6441 with 70.6% accuracy. In binary evaluation (Normal vs. Anomaly), App+AppFT Fusion achieved a macro F1-score of 0.8238 with 82.6% accuracy. Generalization testing on a public fall detection dataset (GMDCSA + LE2I) yielded an F1-score of 0.8904 with 94.7% accuracy. All fusion configurations satisfy real-time requirements at ≥25 FPS inference speed.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi Anomali, Deep Learning, LSTM, Temporal Attention, Late Fusion, ResNet-50, MediaPipe, CCTV, Anomaly Detection, Deep Learning, LSTM, Temporal Attention, Late Fusion, ResNet-50, MediaPipe, CCTV.
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Giobert Boyhosea Panjaitan
Date Deposited:	25 Jul 2026 08:53
Last Modified:	25 Jul 2026 08:53
URI:	http://repository.its.ac.id/id/eprint/138053

Actions (login required)

View Item