Template-Type: ReDIF-Article 1.0
Author-Name:Ntambara Etienne
Author-Workplace-Name:ICT Department, Rwanda Polytechnic Huye College, Huye, Rwanda.
Author-Name:Rukundo Simeon
Author-Workplace-Name:Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore.
Author-Name: Nkurunziza Egide
Author-Workplace-Name:ICT Department, Rwanda Polytechnic Huye College, Huye, Rwanda.
Author-Name: Ingabire Daria
Author-Workplace-Name:ICT Department, Rwanda Polytechnic Huye College, Huye, Rwanda.
Title:Human Activity Classification for Electricity Infrastructure Protection in Rwanda
Abstract:The theft and vandalism of REG electricity infrastructure can disrupt essential services, raise maintenance costs, and pose safety hazards. However, no prior work has established a leakage-controlled, balanced image classification benchmark for activity cues relevant to this operational context. This study developed and tested an AI-based human activity recognition system as an early-warning tool to protect Rwanda Energy Group (REG) electricity infrastructure. We assembled two public Kaggle image datasets, Wall Climbing ALEE and Climbing V2, into six classes: climbing, cutting, normal, opening, suspicious, and vendors. The final dataset included 23,647 images: 19,554 for training, 1,814 for validation, and 2,279 for testing. Oversampling during training reduced the imbalance ratio from 13.751 to 1.000, while maintaining validation and test distributions. We fine-tuned a pretrained YOLO11m model with 224×224 pixel inputs, applying augmentation, AdamW optimization, dropout, cosine learning-rate scheduling, and early stopping. On the untouched test set, the model achieved 99.74% top-1 accuracy, 99.12% balanced accuracy, 99.40% macro F1-score, 0.9998 macro ROC-AUC, and 0.9971 macro average precision; it misclassified only six images out of 2,279. Suspicious activity detection remains a primary challenge, with 95.95% recall and 97.93% F1-score. GPU benchmarking on an NVIDIA Tesla T4 showed a mean latency of 6.64 ms and throughput of about 150.50 images/sec. These results support deploying the model as a high-performance visual perception layer, though operational use will require temporal modeling, sensor fusion, field validation, human oversight, and privacy safeguards.
Keywords:Human activity recognition, CCTV, YOLO11, Electricity infrastructure security, Early warning, Rwanda Energy Group, REG
Journal: Journal of Scientific Reports
Pages:103-127
Volume: 15
Issue: 1
Year: 2026
DOI:10.58970/JSR.1242
File-URL: https://ijsab.com/wp-content/uploads/1242.pdf
File-Format: Application/pdf
File-URL: https://www.ijsab.com/jsr-volume-15-issue-1/9210
File-Format: text/html
Handle: RePEc:aif:report:v:15:y:2026:i:1:p:103-127