Kamera Görüntülerinde Sigara Tespiti için SAHI Destekli YOLO Tabanlı Model Seçimi ve Sabah Saatleri İçin Zaman Bazlı Senaryo Analizi

Abstract

Bu çalışmada, kamera görüntülerinden sigara tespitine yönelik olarak senaryo temelli bir veri düzeni, çok aşamalı model seçimi ve SAHI (Slicing Aided Hyper Inference) destekli çıkarım yaklaşımı birlikte değerlendirilmiştir. Çalışmanın çıkış noktası, sigaranın görüntü içinde çoğu zaman küçük ve ince bir nesne olarak yer alması, buna ek olarak çekim ortamının çok sayıda pencere, koruma demiri, demir yüzeylerin ortama yakın renkleri ve gün içinde değişen ışık koşulları nedeniyle zorlayıcı bir yapıda olmasıdır. Bu nedenle yalnızca tek bir model sonucu raporlamak yerine, önce temel YOLO tabanlı karşılaştırma yapılmış, daha sonra küçük nesne görünürlüğünü artırmak amacıyla SAHI ile ikinci bir karşılaştırma yürütülmüş, ardından seçilen aday modeller daha uzun çevrimlerde yeniden değerlendirilmiştir. Veri yapısı kamuya açık veri ile araştırmacı tarafından oluşturulan yerel verinin birlikte kullanıldığı bir düzende kurulmuştur. Yerel veri tarafında 4.060 sigara içeren görüntü, 3.840 yalnızca ortamı gösteren görüntü, yanlış alarm üretmeye yatkın 1.954 görüntü, 2.178 farklı kişi odaklı pozitif görüntü ve 1.542 kişi odaklı yanlış alarm örneği kullanılmıştır. Çalışma senaryo bazlı tasarlanmış, zaman ekseninde ise sabah saatleri 09.00–14.00 aralığı ana değerlendirme bölgesi olarak belirlenmiştir. Temel YOLO karşılaştırmasında YOLO11m ve YOLO8m öne çıkmış; SAHI destekli ikinci aşamada da YOLO11m en yüksek başarıyı sürdürmüştür. Daha uzun çevrimli değerlendirmeler ve Optuna ile yapılan ayar iyileştirmesi sonunda nihai aday yine YOLO11m olmuştur. Saat bazlı sonuçlarda 11. saat en güçlü, 14. saat ise en zayıf aralık olarak belirlenmiştir. Örnek çerçeveler üzerinden yapılan kutu düzeyi doğrulama, performanstaki bozulmanın yalnızca kutu yerleşiminden değil, bazı saatlerde adayın yeterince bulunamamasından da kaynaklandığını göstermiştir. Ayrıca yanlış alarm incelemesi, ayar sıkılaştırması ile belirgin bir iyileşme elde edildiğini, ancak yanlış alarm probleminin tamamen ortadan kalkmadığını ortaya koymuştur. Elde edilen bulgular, sigara tespitinin yalnızca bir nesne bulma problemi olarak değil, çevresel koşullardan ve zaman bazlı değişkenlikten etkilenen bütüncül bir izleme problemi olarak ele alınması gerektiğini göstermektedir.

Keywords

• Sigara tespiti
• küçük nesne tespiti
• SAHI
• YOLO
• kamera görüntüsü
• yanlış alarm
• zaman bazlı analiz

Supporting Institution

İstanbul Topkapı Üniversitesi

Ethical Statement

Kamera Görüntüleri Üzerinden Sigara Tespit Sistemi (E-49846378-050.04-2600003527)

SAHI-Assisted YOLO-Based Model Selection for Cigarette Detection in Camera Imagery and Time-Based Scenario Analysis for Morning Hours

Abstract

In this study, a scenario-based data structure, a multi-stage model selection process, and a SAHI (Slicing Aided Hyper Inference)-supported inference approach were evaluated together for cigarette detection from camera images. The starting point of the study is that cigarettes often appear as small and thin objects in the image, and, in addition, the shooting environment has a challenging structure due to many windows, protective bars, metal surfaces with colors close to the environment, and changing lighting conditions throughout the day. Therefore, instead of reporting only a single model result, a basic YOLO-based comparison was first conducted, then a second comparison was carried out with SAHI in order to increase small-object visibility, and afterward the selected candidate models were re-evaluated in longer training cycles. The data structure was established in a setting where public data and local data created by the researcher were used together. On the local data side, 4,060 images containing cigarettes, 3,840 images showing only the environment, 1,954 images prone to producing false alarms, 2,178 positive images focused on different individuals, and 1,542 person-focused false alarm examples were used. The study was designed on a scenario basis, and on the time axis, the morning period between 09.00 and 14.00 was determined as the main evaluation region. In the basic YOLO comparison, YOLO11m and YOLO8m stood out; in the SAHI-supported second stage, YOLO11m also maintained the highest performance. After longer-cycle evaluations and hyperparameter improvement performed with Optuna, the final candidate was again YOLO11m. In the hour-based results, the 11th hour was identified as the strongest interval, while the 14th hour was identified as the weakest interval. Box-level validation performed on sample frames showed that the performance degradation was not only caused by box localization, but also by the insufficient detection of candidates in some hours. In addition, the false alarm analysis revealed that a clear improvement was achieved through stricter settings, but that the false alarm problem was not completely eliminated. The findings obtained show that cigarette detection should be considered not only as an object detection problem, but also as a holistic monitoring problem affected by environmental conditions and time-based variability.

Keywords

• Cigarette detection
• small object detection
• SAHI
• YOLO
• camera imagery
• false positives
• time-based analysis

Supporting Institution

İstanbul Topkapı Üniversity

Ethical Statement

Kamera Görüntüleri Üzerinden Sigara Tespit Sistemi (E-49846378-050.04-2600003527)

References

1. [1] A. Khan, S. Khan, B. Hassan, and Z. Zheng, "CNN-Based Smoker Classification and Detection in Smart City Application," Sensors, vol. 22, no. 3, p. 892, 2022, doi: 10.3390/s22030892.
2. [2] E. C. Santiago, E. M. Reyes, M. L. Tria, J. V. Obmerga, and R. C. Reyes, "Deep Convolutional Neural Network for Detection of Cigarette Smokers in Public Places: A Low Sample Size Training Data Approach," in 2022 International Conference on Decision Aid Sciences and Applications (DASA), 2022, pp. 1276-1280, doi: 10.1109/DASA54658.2022.9765180.
3. [3] A. Pundhir, D. Verma, P. Kumar, and B. Raman, "Region Extraction Based Approach for Cigarette Usage Classification Using Deep Learning," in Computer Vision and Image Processing. CVIP 2021. Communications in Computer and Information Science, vol. 1568, Springer, 2022, doi: 10.1007/978-3-031-11349-9_33.
4. [4] Z. Wang, L. Lei, and P. Shi, "Smoking Behavior Detection Algorithm Based on YOLOv8-MNC," Frontiers in Computational Neuroscience, vol. 17, p. 1243779, 2023, doi: 10.3389/fncom.2023.1243779.
5. [5] J. Chong, "An Intelligent Detection Approach for Smoking Behavior," International Journal of Cognitive Informatics and Natural Intelligence, vol. 17, no. 1, pp. 1-18, 2023, doi: 10.4018/IJCINI.324115.
6. [6] Y. Bakhytov and C. Öz, "Cigarette Detection in Images Based on YOLOv8," Sakarya University Journal of Computer and Information Sciences, vol. 7, no. 2, pp. 253-263, 2024, doi: 10.35377/saucis.1461268.
7. [7] Y. Li, H. Zhou, J. Feng, X. Li, X. Xu, P. Hou, and X. Hu, "An Improved Smoking Behavior Detection Algorithm via Incorporating an Interference Information Filtering Network," Engineering Applications of Artificial Intelligence, vol. 136, p. 109050, 2024, doi: 10.1016/j.engappai.2024.109050.
8. [8] Y. Zhu, "Smoking Detection Model Based on Improved YOLOv8-s," in Proceedings of the International Conference on Intelligent Computing (ICIC), 2024, doi: 10.65286/icic.v20i1.92233.

9. [9] A. Khan, M. A. M. Elhassan, S. Khan, and H. Deng, "Deep Learning-Based Smoker Classification and Detection: An Overview and Evaluation," Expert Systems with Applications, vol. 267, p. 126208, 2025, doi: 10.1016/j.eswa.2024.126208.
10. [10] G. Gkioxari, R. Girshick, P. Dollár, and K. He, "Detecting and Recognizing Human-Object Interactions," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 8359-8367.
11. [11] H.-B. Ling and D. Huang, "Single-Image Smoker Detection by Human-Object Interaction with Post-Refinement," in Neural Information Processing. ICONIP 2021. Lecture Notes in Computer Science, Springer, 2021, pp. 181-192, doi: 10.1007/978-3-030-92270-2_16.
12. [12] H.-B. Ling, D. Huang, J. Cui, and C.-D. Wang, "HOLT-Net: Detecting Smokers via Human-Object Interaction with Lite Transformer Network," Engineering Applications of Artificial Intelligence, vol. 126, p. 106919, 2023, doi: 10.1016/j.engappai.2023.106919.
13. [13] W. Zhang and A. L. De Ocampo, "Detection of Smoking Behavior by Identifying Cigarettes and Key Action Features Using YOLO Algorithm," in Proceedings of SPIE, vol. 13688, 2025, Art. no. 1368804, doi: 10.1117/12.3075841.
14. [14] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie, "Feature Pyramid Networks for Object Detection," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 2117-2125.
15. [15] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, "Focal Loss for Dense Object Detection," in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2980-2988.
16. [16] B. Mirzaei, H. Nezamabadi-Pour, A. Raoof, and R. Derakhshani, "Small Object Detection and Tracking: A Comprehensive Review," Sensors, vol. 23, no. 15, p. 6887, 2023, doi: 10.3390/s23156887.
17. [17] M. Nikouei, B. Baroutian, S. Nabavi, F. Taraghi, A. Aghaei, A. Sajedi, and M. E. Moghaddam, "Small Object Detection: A Comprehensive Survey on Challenges, Techniques and Real-World Applications," arXiv preprint arXiv:2503.20516, 2025. [Online]. Available: https://arxiv.org/abs/2503.20516.
18. [18] Z. Yuan, J. Gong, B. Guo, C. Wang, N. Liao, J. Song, and Q. Wu, "Small Object Detection in UAV Remote Sensing Images Based on Intra-Group Multi-Scale Fusion Attention and Adaptive Weighted Feature Fusion Mechanism," Remote Sensing, vol. 16, no. 22, p. 4265, 2024, doi: 10.3390/rs16224265.
19. [19] F. C. Akyon, S. O. Altinuc, and A. Temizel, "Slicing Aided Hyper Inference and Fine-tuning for Small Object Detection," arXiv preprint arXiv:2202.06934, 2022. [Çevrimiçi]. Erişim: https://arxiv.org/abs/2202.06934.
20. [20] T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, "Optuna: A Next-generation Hyperparameter Optimization Framework," in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2019, pp. 2623-2631, doi: 10.1145/3292500.3330701.
21. [21] J. Redmon and A. Farhadi, "YOLO9000: Better, Faster, Stronger," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 7263-7271, doi: 10.1109/CVPR.2017.690.
22. [22] Ultralytics, "YOLO11 Documentation," Ultralytics Official Documentation, 2024. [Çevrimiçi]. Erişim: https://docs.ultralytics.com/models/yolo11/.
23. [23] Takoyati, "Cigarette Vape Detection Computer Vision Dataset", Roboflow Universe, 2025. [Çevrimiçi]. Erişim: https://universe.roboflow.com/takoyati/cigarette-vape-detection/dataset/1.
24. [24] Richie Lab, "Smoking Computer Vision Model," Roboflow Universe, 2025. [Çevrimiçi]. Erişim: https://universe.roboflow.com/richie-lab/smoking-tasfx/dataset/1.
25. [25] Camera AI, "Smoking Computer Vision Dataset", Roboflow Universe, 2025. [Çevrimiçi]. Erişim: https://universe.roboflow.com/camera-ai-qj14x/smoking-hvni4-oaxdv/dataset/1.
26. [26] Cigarette, "Cigarette Computer Vision Model," Roboflow Universe, 2025. [Çevrimiçi]. Erişim: https://universe.roboflow.com/cigarette-c6554/cigarette-ghnlk/dataset/1.
27. [27] "Smoking Detection Public Dataset," 2026. [Çevrimiçi]. Erişim: https://anonymous.4open.science/r/Smoking-Detection-Public-Dataset-6E8C/
28. [28] M. Hamidaoui, M. Z. Talhaoui, M. Li, M. A. Midoun, S. Haouassi, D. E. Mekkaoui, A. Smaili, A. Cherraf, and F. Z. Benyoub, "Survey of Autonomous Vehicles' Collision Avoidance Algorithms," Sensors, vol. 25, no. 2, p. 395, 2025, doi: 10.3390/s25020395.