10023 research outputs found
Sort by
Optimization of Vehicle Detection at Intersections Using the YOLOv5 Model
Get PDFThis study aims to analyze and evaluate the performance of the YOLOv5 model in detecting vehicles at intersections to optimize traffic flow. The methods used in this research include training the YOLOv5 model with traffic datasets collected from various intersections and optimizing hyperparameters to achieve the best detection accuracy. The study results show that the optimized YOLOv5 model can detect multiple types of vehicles with high accuracy. The model achieved a detection accuracy of 85.47% for trucks, 87.12% for pedestrians, 86.54% for buses, 77.20% for cars, 80.48% for motorcycles, and 78.80% for bicycles. Significant improvements in detection performance were achieved compared to the default model. This research concludes that the optimization of the YOLOv5 model is effective in improving vehicle detection accuracy at intersections. Implementing this optimized model can significantly contribute to traffic management, reduce congestion, and improve road safety. It is expected that the implementation of this technology can be more widely applied for more efficient traffic management in various major cities
Optimizing South Kalimantan Food Image Classification Through CNN Fine-Tuning
Get PDFSouth Kalimantan's rich culinary heritage encompasses numerous traditional dishes that remain unfamiliar to visitors and digital platforms. While Convolutional Neural Networks (CNNs) have demonstrated remarkable success in image classification tasks, their application to regional cuisine faces unique challenges, particularly when dealing with limited datasets and visually similar dishes. This study addresses these challenges by evaluating and optimizing two pre-trained CNN architectures—EfficientNetB0 and InceptionV3—for South Kalimantan food classification. Using a custom dataset of 1,000 images spanning 10 traditional dishes, we investigated various fine-tuning strategies to maximize classification accuracy. Our results show that EfficientNetB0, with 30 fine-tuned layers, achieves the highest accuracy at 94.50%, while InceptionV3 reaches 92.00% accuracy with 40 layers fine-tuned. These findings suggest that EfficientNetB0 is particularly effective for classifying regional foods with limited data, outperforming InceptionV3 in this context. This study provides a framework for efficiently applying CNN models to small, specialized datasets, contributing to both the digital preservation of South Kalimantan’s culinary heritage and advancements in regional food classification. This research also opens the way for further research that can be applied to other less documented regional cuisines. The framework presented can be used as a reference for developing automated classification systems in a broader cultural context, thus enriching the digital documentation of traditional cuisines and preserving the culinary diversity of the archipelago for future generations
