Publication Search

This study proposes a computer vision-based system for automatically verifying the use of Personal Protective Equipment (PPE) in industrial environments. The system integrates the YOLOv8 object detection model, OpenCV for image processing, and a State Machine mechanism to manage the verification workflow. The verification process begins with employee identification through ID card scanning, followed by real-time detection of the head, safety helmet, and face mask using a camera, before generating a final PASS or FAIL decision. System evaluation was conducted using 250 testing scenarios to assess both model and system performance. The results show that the YOLOv8 model achieved an mAP@50 of 93.9%, while the overall verification system obtained 98.4% accuracy, 98.4% precision, 100% recall, and a 99.2% F1-score. The implementation of the State Machine contributed to a more stable and consistent verification process by ensuring that each inspection stage was executed in the correct sequence. These findings demonstrate that the proposed system can effectively support automated PPE compliance monitoring and has the potential to enhance occupational safety management in industrial workplaces. 

Modern food supply chains, particularly those involving essential commodities like rice, often suffer from major challenges such as product fraud, inefficient record-keeping, and a lack of consumer trust. Traditional centralized systems are prone to data tampering, limited transparency, and poor traceability, making it difficult to verify the authenticity and origin of goods. To address these issues, our research introduces TraceRoot, a blockchain-based traceability framework designed to enhance transparency, accountability, and trust in agricultural supply chains.TraceRoot leverages the immutability and decentralization of blockchain technology to maintain a secure, distributed ledger that records every transaction and movement of goods across the supply chain. Each stakeholder including farmers, distributors, retailers, and consumers has role-based access to authenticated data through a user-friendly interface. The framework integrates smart contracts to automate transactions and digital signatures to verify the integrity of the data being uploaded, minimizing the risk of human error or manipulation

Evaluating food portions and types in the Free Nutritious Meal (MBG) program is generally still performed manually, which is time-consuming and potentially subjective. This study aims to develop an automated deep learning-based system to efficiently detect food types and estimate their nutritional value. The method used is a quantitative experiment integrating the YOLOv11m architecture for real-time object detection and the Google Gemini 2.5 Flash Large Language Model (LLM) for contextual nutritional estimation reasoning. The model training utilized a dataset of 2,630 food tray images categorized into five classes (fruit, side dish, staple food, vegetable, milk) that had undergone an augmentation process. The results showed that the YOLOv11m model achieved excellent performance with a mean Average Precision (mAP@0.5) of 0.9727 and the highest F1-score of 0.9522 at a confidence threshold of 0.1. Furthermore, validation of the LLM integration demonstrated a high prediction agreement rate of 85%. In conclusion, the combination of the YOLOv11m algorithm and LLM reasoning is capable of detecting and validating nutritional classification quickly and precisely, showing strong potential as an objective nutritional evaluation monitoring solution for large-scale MBG program implementation.

Malware represents an evolving cybersecurity threat that demands more effective detection methods. Conventional signature-based detection systems have limitations in identifying new variants, driving the development of deep learning-based approaches. This research implements and evaluates four variants of the YOLOv11 algorithm (n, s, m, l) for malware classification based on visual image representation. The dataset consists of 22,056 malware and benign images, divided into 70% training, 15% validation, and 15% testing across 8 classes (adware, backdoor, benign, downloader, spyware, trojan, virus, worm). Each model was trained for 100 epochs with batch size 32 using Google Colab with GPU support. Results demonstrate that all variants achieve high accuracy (97.8%-98.1%) with YOLOv11m as the best performer (98.1%). YOLOv11n offers optimal balance between accuracy (97.9%) and efficiency (1.5M parameters, 0.3 ms/img inference) ideal for real-time applications. This research surpasses previous methods such as K-NN (97.18%) and hybrid CNN (96.55%) with superior inference speed (0.3-0.9 ms/img vs tens to hundreds of ms/img), proving the effectiveness of YOLOv11 for fast, accurate, and scalable malware detection.

Abstract: Detecting chili leaf diseases remains challenging due to the non-uniform manifestation of symptoms, local discoloration, small lesion regions, and visual similarity between disease patterns and natural leaf background variations. Although YOLO-based detectors provide favorable computational efficiency, lightweight variants often struggle to distinguish subtle lesion characteristics, while conventional attention mechanisms such as CBAM primarily rely on global feature aggregation and may overlook regional activation variability. To address these limitations, this study proposes a YOLOv9s-based detection framework integrated with a Region-Dispersion Channel Spatial Attention (RDCSA) module. The proposed module incorporates regional dispersion statistics, namely mean, standard deviation, and range, as channel descriptors to capture inter-region feature variability before applying spatial attention refinement. Experiments were conducted on the COLD dataset containing 532 original images from five chili leaf condition categories using a split-before-augmentation protocol to ensure objective evaluation. RDCSA was integrated at the P5 feature level and evaluated through attention placement analysis, component-wise ablation, sensitivity analysis, stability assessment, and comparison with modern attention mechanisms. The proposed YOLOv9s + RDCSA model achieved an mAP@50 of 0.894, mAP@50–95 of 0.773, precision of 0.858, recall of 0.861, and an F1-score of 0.859 with only a marginal increase in model parameters. The results suggest that regional dispersion-based attention improves feature discrimination while preserving computational efficiency, particularly for disease symptoms characterized by heterogeneous spatial patterns. Nevertheless, performance remains influenced by visually ambiguous symptom categories, indicating that further validation across multiple datasets and field conditions is required. Overall, the proposed RDCSA module enhances detection capability without substantially increasing computational overhead, making it a promising attention mechanism for lightweight plant disease detection systems.

Bullying is a serious problem that can disrupt the learning process and mental development of students, including in Islamic boarding schools. Early detection of bullying is essential to creating a safe and conducive learning environment. This study aims to apply the You Only Look Once (YOLO) algorithm to automatically detect bullying through video recordings in the environment of the SMK Skill Village Islamic School Business Boarding School. The method used involves collecting a video dataset representing various types of bullying behavior, labeling the data, and training an object detection model using the YOLOv5 algorithm. The developed system is capable of detecting and classifying bullying behavior in real- time with detection accuracy reaching [accuracy value if known]. The implementation of this system is expected to assist school authorities and boarding school administrators in monitoring, preventing, and addressing bullying incidents more quickly and effectively, while also serving as an initial step in leveraging artificial intelligence technology to create a safer and more comfortable educational environment.

This research is motivated by the problems in manual attendance systems at schools, which remain vulnerable to fraud, time-consuming, and inefficient. The expected solution is to develop an automated attendance system based on face recognition that can operate in realtime with high accuracy. The research object is vocational high school students, with the applied method implementing the YOLO v10 algorithm for face detection, followed by the face_recognition library for identification. The instruments used include an Imou CCTV camera as the input device, a mid-range laptop as the hardware platform, and Python with SQLite as the software environment for data processing and attendance storage. The results show that the developed system achieved an average face detection accuracy of 96% under normal lighting and 91% under low lighting, with an average processing speed of 27 FPS. The implementation of an anti-duplication feature also ensured data validity by allowing each student to be recorded only once per day. In conclusion, the use of YOLO v10 in face-based attendance proved to be effective, efficient, and capable of reducing fraud. The implication of this study is that the system can be applied in both Islamic boarding schools and general schools as a modernization of attendance systems, with a recommendation for further development through web-based application and cloud database integration.

Tanaman padi merupakan salah satu komoditas pangan utama di dunia yang memiliki peran penting dalam menjaga ketahanan pangan global. Namun, produktivitas tanaman padi seringkali terganggu oleh berbagai penyakit daun yang dapat menurunkan kualitas dan hasil panen secara signifikan. Deteksi penyakit secara dini menjadi langkah penting untuk mengendalikan penyebaran penyakit tersebut. Metode konvensional yang mengandalkan pengamatan manual memiliki keterbatasan dalam hal kecepatan, akurasi, serta ketergantungan terhadap keahlian manusia. Oleh karena itu, diperlukan suatu sistem deteksi otomatis yang mampu mengidentifikasi penyakit secara cepat dan akurat. Penelitian ini bertujuan untuk mengembangkan model deteksi penyakit daun padi menggunakan metode YOLO (You Only Look Once) berbasis Deep Learning. Dataset yang digunakan diperoleh dari Kaggle yang terdiri dari citra daun padi dengan anotasi bounding box. Proses penelitian meliputi tahap pengumpulan data, preprocessing, pelatihan model, serta evaluasi performa. Model dilatih menggunakan optimizer Stochastic Gradient Descent (SGD) dengan parameter tertentu untuk memperoleh hasil yang optimal. Evaluasi kinerja model dilakukan menggunakan metrik mean Average Precision (mAP), precision, dan recall. Hasil penelitian menunjukkan bahwa model YOLO mampu mendeteksi penyakit daun padi dengan tingkat akurasi yang tinggi serta waktu deteksi yang relatif cepat. Hal ini menunjukkan bahwa metode YOLO efektif untuk diterapkan dalam sistem pertanian cerdas. Penelitian ini diharapkan dapat memberikan kontribusi dalam pengembangan teknologi deteksi penyakit tanaman berbasis kecerdasan buatan.

Perkembangan teknologi artificial intelligence, khususnya dalam bidang computer vision dan deep learning, telah mendorong lahirnya berbagai sistem otomatis dalam analisis citra dan video. Salah satu implementasi yang banyak digunakan adalah deteksi objek secara real-time menggunakan algoritma YOLO (You Only Look Once). Di sisi lain, penggunaan rokok elektronik atau vape semakin meningkat, terutama di lingkungan pendidikan, sehingga diperlukan sistem pengawasan otomatis yang mampu mendeteksi aktivitas tersebut secara akurat. Penelitian ini bertujuan untuk mengembangkan sistem deteksi aktivitas vape berbasis YOLOv8 pada citra dan video. Metode yang digunakan dalam penelitian ini adalah pendekatan deep learning workflow yang meliputi pengumpulan dataset, pra-pemrosesan, pelatihan model, serta evaluasi performa. Dataset yang digunakan dikembangkan secara mandiri dengan jumlah awal 1000 gambar, kemudian ditingkatkan menjadi 2000 gambar melalui proses augmentasi menggunakan Roboflow. Proses pelatihan dilakukan dengan parameter tertentu, seperti 100 epoch dan ukuran citra 640×640 piksel. Evaluasi model dilakukan menggunakan metrik precision, recall, mean Average Precision (mAP), serta kecepatan deteksi (frame per second). Hasil penelitian menunjukkan bahwa model YOLOv8 mampu mendeteksi aktivitas vape dengan performa yang cukup baik. Model mencapai nilai precision sebesar 0,91, recall sebesar 0,88, mAP@0.5 sebesar 0,92, serta mAP@0.5:0.95 sebesar 0,76. Selain itu, sistem mampu bekerja secara real-time dengan kecepatan deteksi mencapai 28 FPS. Hasil pengujian juga menunjukkan bahwa model mampu mengenali objek vape pada berbagai kondisi lingkungan, meskipun masih mengalami kendala pada objek berukuran kecil dan kondisi occlusion. Dengan demikian, penelitian ini membuktikan bahwa algoritma YOLOv8 efektif digunakan untuk mendeteksi aktivitas vape pada citra dan video, serta memiliki potensi untuk diterapkan dalam sistem pengawasan berbasis CCTV secara real-time.

Penelitian ini bertujuan untuk menganalisis dan membandingkan kinerja model YOLOv11 dan MobileNetV3 dalam mengklasifikasikan varietas padi berdasarkan gambar digital. Kumpulan data terdiri dari 10.000 gambar yang mewakili lima varietas padi, yaitu Arborio, Basmati, Ipsala, Jasmine, dan Karacadag. Kumpulan data tersebut dibagi menjadi set pelatihan dan set pengujian, dengan proses pelatihan dilakukan selama 100 epoch. Hasil evaluasi menunjukkan bahwa MobileNetV3 mencapai akurasi klasifikasi sebesar 100%, sedangkan YOLOv11 memperoleh akurasi sebesar 99,8%. Meskipun akurasinya sedikit lebih rendah, YOLOv11 menunjukkan kinerja yang stabil dengan kesalahan klasifikasi yang minimal. Analisis matriks kebingungan menunjukkan bahwa sebagian besar prediksi masuk ke dalam kelas yang benar, dengan hanya sedikit kesalahan yang terjadi pada kategori yang secara visual mirip. Temuan ini menunjukkan bahwa kedua model tersebut sangat efektif untuk tugas klasifikasi varietas padi. Namun, evaluasi lebih lanjut menggunakan dataset yang lebih kompleks diperlukan untuk memastikan ketahanan dan generalisasi model dalam skenario dunia nyata.

Object detection plays a crucial role in intelligent transportation systems, particularly for outdoor traffic monitoring applications that require accurate and real-time performance under limited computational resources. Recent developments in YOLO-based architectures have introduced multiple model variants; however, their practical performance under constrained training conditions remains insufficiently explored. This study presents a comparative evaluation of YOLOv5, YOLOv7, and YOLOv8 for outdoor traffic object detection using a real-world dataset and identical experimental settings. The main objective of this research is to analyze the robustness and detection quality of different YOLO variants when trained with a limited number of epochs, reflecting practical deployment scenarios. All models were trained and evaluated using the same dataset, preprocessing pipeline, and hardware configuration to ensure a fair comparison. Performance evaluation was conducted using multiple metrics, including precision, recall, mAP@50, Precision–Recall curves, area under the curve (AUC), and peak F1-score. Experimental results indicate that YOLOv5 outperformed YOLOv7 and YOLOv8 in terms of overall detection stability and robustness. The merged Precision–Recall analysis shows that YOLOv5 achieved a higher effective AUC and superior mAP@50, reflecting better global detection performance. In addition, YOLOv5 exhibited a higher peak F1-score, indicating a more balanced trade-off between precision and recall. In contrast, YOLOv7 and YOLOv8 showed performance degradation under limited training conditions despite their more advanced architectures. These findings suggest that YOLOv5 remains a reliable and efficient solution for outdoor traffic object detection, particularly in resource-constrained environments. The study highlights the importance of comprehensive evaluation metrics and practical experimental settings when selecting object detection models for real-world applications.

Object detection plays a crucial role in intelligent transportation systems, particularly for outdoor traffic monitoring applications that require accurate and real-time performance under limited computational resources. Recent developments in YOLO-based architectures have introduced multiple model variants; however, their practical performance under constrained training conditions remains insufficiently explored. This study presents a comparative evaluation of YOLOv5, YOLOv7, and YOLOv8 for outdoor traffic object detection using a real-world dataset and identical experimental settings. The main objective of this research is to analyze the robustness and detection quality of different YOLO variants when trained with a limited number of epochs, reflecting practical deployment scenarios. All models were trained and evaluated using the same dataset, preprocessing pipeline, and hardware configuration to ensure a fair comparison. Performance evaluation was conducted using multiple metrics, including precision, recall, mAP@50, Precision–Recall curves, area under the curve (AUC), and peak F1-score. Experimental results indicate that YOLOv5 outperformed YOLOv7 and YOLOv8 in terms of overall detection stability and robustness. The merged Precision–Recall analysis shows that YOLOv5 achieved a higher effective AUC and superior mAP@50, reflecting better global detection performance. In addition, YOLOv5 exhibited a higher peak F1-score, indicating a more balanced trade-off between precision and recall. In contrast, YOLOv7 and YOLOv8 showed performance degradation under limited training conditions despite their more advanced architectures. These findings suggest that YOLOv5 remains a reliable and efficient solution for outdoor traffic object detection, particularly in resource-constrained environments. The study highlights the importance of comprehensive evaluation metrics and practical experimental settings when selecting object detection models for real-world applications.

Indramayu mango is a seasonal fruit that is highly favored due to its delicious taste and high nutritional content. However, high mango production is often not supported by adequate post-harvest facilities, particularly in terms of fruit ripeness classification. Currently, mango ripeness classification is still performed manually, which tends to be subjective and inconsistent. To address this issue, this study proposes a ripeness detection system for Indramayu mangoes by integrating the TGS2602 gas sensor and the YOLOv11 algorithm based on image processing. The TGS2602 sensor is used to detect ethylene gas emitted by ripe mangoes, while YOLOv11 is employed for visual image analysis of the fruit. This study aims to evaluate the system’s performance in classifying ripe and unripe mangoes, as well as analyze the integration between the gas sensor and the object detection model. The test results show that the TGS2602 sensor can detect increased ethylene gas concentration in ripe mangoes, while YOLOv11 demonstrates high accuracy in detecting mangoes based on visual images, with precision and recall close to 1.0. The system was also tested under various lighting conditions, including dark environments, and still performed well, although with a slight decrease in accuracy under low-light conditions.

Background: Accurate detection of renal cysts on CT urography requires high diagnostic precision, while manual interpretation by radiologists is susceptible to inter-observer variability and potential delays in clinical decision-making. These challenges underscore the need for a reliable automated detection system to support radiological assessment. Objective: This study aims to develop and evaluate the performance of the Neo-ZasAI application based on the YOLOv8 algorithm for the automatic identification of renal cysts. Methods: Employing a Research and Development design using the ADDIE model, the study encompassed needs analysis, model design, software development, system implementation using 200 CT urography images, and diagnostic performance evaluation. Classification results generated by Neo-ZasAI were compared with radiologist readings through confusion matrix analysis and ROC–AUC assessment. Results: The findings indicate that Neo-ZasAI achieved an accuracy of 97,5%, sensitivity of 96%, specificity of 99%, positive predictive value of 98,9%, and negative predictive value of 96,1%. The ROC analysis yielded an AUC of 0.988 (p < 0.001), demonstrating excellent discriminative capability and high concordance with radiologist interpretations as the diagnostic gold standard. Conclusion: These results suggest that Neo-ZasAI is capable of performing rapid, consistent, and accurate renal cyst detection and is thus feasible for implementation as a clinical decision support system in radiology, with potential integration into PACS workflows and further development to enhance model generalizability.

This study aims to develop an Indonesian traffic sign detection system using a transfer learning approach to improve road safety and traffic efficiency. The dataset was obtained from Kaggle and consists of 2,100 images across 21 traffic sign classes. The research stages include data collection, preprocessing to reduce noise and normalize image brightness, object detection using YOLOv5, and classification based on transfer learning with ResNet, VGG-16, and MobileNet architectures. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics. Experimental results indicate that the YOLOv5 model is capable of detecting traffic sign objects; however, the classification performance remains relatively low, with a mean Average Precision (mAP) value of 0.17. These findings suggest that further optimization is required in data preprocessing, dataset quality, and model parameter tuning to achieve better performance. This study demonstrates that transfer learning has significant potential for developing computer vision-based traffic sign detection systems, although further improvements are necessary to ensure robustness under real-world Indonesian traffic conditions.

Akumulasi limbah yang tidak terkelola secara efektif menimbulkan ancaman serius bagi keberlanjutan lingkungan dan kesehatan masyarakat global, khususnya di kawasan urban padat penduduk. Meskipun teknologi pemilahan sampah otomatis telah berkembang, sistem eksisting sering kali terkendala efisiensi energi yang rendah dan isu higienitas akibat ketergantungan pada kontak fisik atau penggunaan sistem kamera yang aktif terus-menerus (always-on). Penelitian ini bertujuan merancang Sistem Pemilah Sampah Cerdas Hibrida yang mengintegrasikan teknologi visi komputer algoritma YOLOv8 Nano dengan sensor ultrasonik HC-SR04 untuk mewujudkan solusi pemilahan nirkontak (touchless). Menggunakan metode Research and Development (R&D), sistem ini menawarkan mekanisme manajemen daya efisien, di mana sensor ultrasonik bertindak sebagai pemicu jarak dekat (<10 cm) yang secara otomatis membuka pintu sampah dan mengaktifkan kamera akuisisi citra hanya ketika objek terdeteksi. Model YOLOv8 dilatih menggunakan dataset sebanyak 25.077 citra yang bersumber dari repositori publik Kaggle, dan berhasil mencapai performa Presisi sebesar 80,6% serta mAP@50 sebesar 61,7%. Berdasarkan pengujian real-time, sistem mampu mengklasifikasikan dan menginstruksikan aktuator servo untuk memisahkan sampah ke wadah organik atau anorganik secara akurat. Dilengkapi modul IoT ESP8266, sistem ini juga memfasilitasi pemantauan kapasitas volume sampah. Kesimpulannya, integrasi cerdas antara pemicu sensor dan deep learning ini tidak hanya memperbaiki akurasi pemilahan, tetapi juga menghadirkan solusi yang lebih higienis, responsif, dan hemat energi sebagai pendukung ekosistem Smart City.

The increasing intensity of traffic object movement in urban areas has not been accompanied by adequate road infrastructure, resulting in traffic congestion, air pollution, and a higher risk of traffic accidents. One of the primary causes of accidents is traffic violations, particularly wrong-way driving behavior. This study develops a video-based automated traffic violation detection system using the YOLOv5 algorithm. A computer vision approach is employed to detect, classify traffic objects, and count wrong-way violations in real time. Due to limited access to real-world traffic violation footage, simulated traffic scenarios are used as testing data. The system is evaluated on four traffic object classes: motorcycles, cars, buses, and trucks. Experimental results demonstrate strong performance, achieving a precision of 90%, a recall of 92%, and an F1-score of 91%, while the traffic object counting accuracy reaches 89%. These findings indicate that the proposed system has significant potential to support traffic analysis and assist authorities in making more effective decisions to reduce congestion and traffic accidents.

Traffic congestion is a persistent challenge in urban areas in Indonesia, where increasing vehicle density creates the need for intelligent traffic monitoring systems. This study aims to develop a real-time vehicle parking system using the YOLOv8 object detection model to provide efficient traffic analysis from live CCTV broadcasts and recorded videos. This study uses a quantitative experimental approach with the implementation of the YOLOv8m model using the Ultralytics library in Python, tested on data collected from CCTV cameras A TCS Dishub Medan and additional footage from mobile devices. Vehicles are detected and counted in two directions up (Up) and down (Down) using virtual detection lines on the video frame. The system performance is evaluated by automatic detection counting with manually recorded ground truth data. The results show that on live CCTV broadcasts, the YOLOv8m model achieves an average precision of 98.96%, a recall of 96.59%, and an F1 score of 97.74% for upstream traffic, while for downstream traffic it achieves 100% precision, 95.64% recall, and an F1 score of 97.730/0. On the other hand, on high-quality recorded videos, all performance metrics achieve 100%, indicating perfect detection accuracy. These findings confirm the effectiveness of YOLOv8 in real-time traffic monitoring, but also indicate that video quality and stream stability affect detection performance. In conclusion, the developed system shows strong potential to support smart city traffic management solutions. Future research should focus on performance optimization under low-resolution live streaming conditions to improve accuracy in practical applications.  

Sudden Infant Death Syndrome (SIDS) is a sudden and unexpected death in infants that is often associated with the prone sleeping position. This study aims to develop an automated monitoring system capable of detecting SIDS risk factors using the YOLOv8 algorithm and to analyze the effect of data augmentation on model performance. The dataset consists of two classes, baby-lying-on-back (supine) and baby-lying-on-stomach (prone), which were processed through model training and evaluation using precision, recall, F1-score, and mAP metrics. The model was trained under two scenarios, without data augmentation and with data augmentation. The results show that the model without augmentation achieved a precision of 90%, recall of 85%, F1-score of 86%, and mAP50 of 93.7%. After applying augmentation, performance improved to a precision of 90%, recall of 87%, F1-score of 88%, and mAP50 of 95.1%. These findings indicate that augmentation increases detection accuracy and enhances model generalization, including robustness against variations in lighting and camera angles. Furthermore, testing with image and video inputs revealed that the non-augmented model exhibited a tendency toward overfitting, particularly in favor of the baby-lying-on-stomach, whereas the augmented model successfully classified both classes accurately. The developed system is also equipped with an alarm feature and early-warning notifications via Telegram to smartphone when a prone position is detected for a certain duration. Overall, the results demonstrate that YOLOv8 with data augmentation is effective for an automated, non-invasive monitoring system for infants, making it suitable for detecting and preventing potential SIDS risk factors.

The use of protective equipment in the form of helmets is an important aspect of ensuring motorcycle rider safety. However, violations of helmet usage still frequently occur and are difficult to monitor continuously. This study proposes a real-time helmet detection system using the YOLOv8 object detection method. The YOLOv8n model was trained using a helmet and no-helmet image dataset that underwent data augmentation to improve the model’s robustness against variations in environmental conditions. The system was implemented using the Python programming language with the support of the Ultralytics and OpenCV libraries. The system input was obtained from a webcam with a resolution of 640×640 pixels, where each video frame was processed in real time to detect the Helmet and No Helmet classes. The system displays bounding boxes and class labels in real time and is equipped with a violation duration calculation mechanism. When a no-helmet condition is detected continuously, the system generates pop-up alerts and automatic notifications via the Telegram application. The experimental results show that the system is capable of detecting helmet usage and no-helmet violations in real time with stable performance. The integration of violation duration calculation helps reduce momentary detection errors and improves the reliability of identifying valid violations