Science and Engineering Connect

Development of an Artificial Intelligence Program with CCTV to Detect Red-Light Violations Based on Traffic Light Colors

Jetsada Kumphong, Nathayu Chawapattanayotha — Tue, 31 Mar 2026 00:00:00 +0700

Background and Objectives: Traffic accidents are one of the leading causes of injury and fatality worldwide, and intersections represent particularly high-risk locations. Among various risky driving behaviors, red-light violations significantly contribute to intersection-related crashes, leading to severe injuries and fatalities. Traditional enforcement mechanisms, such as manual monitoring by traffic police or the use of fixed cameras, are often limited in coverage, accuracy, and efficiency. In recent years, the integration of Artificial Intelligence (AI) with closed-circuit television (CCTV) systems has emerged as a promising solution to enhance law enforcement. AI technology, particularly object detection and image recognition, enables real-time monitoring and automatic detection of traffic signal violations. This innovation is not only a tool for surveillance but also an instrument for behavioral change, as the perceived risk of being caught can deter risky driving practices. The present research was designed to address two primary objectives: (1) to develop and evaluate AI technology for detecting traffic light colors in order to identify red-light violation behaviors, and (2) to study the specific phenomenon of “rolling stop” violations, where drivers reduce speed but fail to completely stop at a red light. The study focuses on six major intersections in Khon Kaen, Thailand, where traffic density and violation rates remain high.

Methodology: The study utilized CCTV footage from six significant intersections in Khon Kaen: Si Than, Bangkok Hospital, Ban Kok, Modin Daeng, Pratu Mueang, and Charoen Sri. A dataset of 557 images was compiled to train and evaluate the AI system. The researchers applied YOLO version 4 (You Only Look Once), a state-of-the-art object detection algorithm capable of identifying objects and classifying them in real time with high accuracy. The model was specifically trained to detect traffic light colors (red, yellow, and green) and correlate them with vehicle behavior at intersections. The performance of the AI system was assessed using a Confusion Matrix, which measures the accuracy of classification by calculating true positives, false positives, true negatives, and false negatives. To analyze red-light violations, a total of 2,041 incidents were documented and coded for statistical analysis. Descriptive statistics provided an overview of violation patterns, while Chi-Square Tests were conducted to examine the relationships between red-light violations and key factors such as vehicle type and time of day. This combination of AI-driven detection and statistical testing ensured both technological reliability and behavioral insights.

Main Result: The results of the study demonstrate the effectiveness of combining AI with CCTV for traffic enforcement. The AI model successfully detected the traffic light color in 94% of the sample images, indicating a high level of accuracy and reliability for law enforcement applications. This accuracy suggests that AI technology can complement or even replace traditional manual monitoring methods, which are often prone to human error and resource limitations. In terms of behavioral patterns, the analysis revealed that vehicle type and time of day were significantly associated with red-light violations (p-value < 0.05). Larger vehicles, such as trucks and buses, showed distinct patterns of non-compliance compared to smaller vehicles. Furthermore, violations were more frequent during peak traffic hours, particularly in the evening, highlighting the role of traffic congestion and driver impatience in risky decision-making. The study also identifies the prevalence of “rolling stop” behavior, where drivers slowed down but failed to come to a complete stop during a red signal. This behavior represents a critical safety concern, as it increases the likelihood of side-impact collisions, especially with pedestrians and motorcyclists who often cross intersections during the red-light phase

Conclusions: The research confirms that AI-powered CCTV systems can substantially enhance the monitoring and enforcement of traffic rules in urban environments. The high detection accuracy of 94% demonstrates the robustness of YOLO v4 as a real-time detection tool for traffic light violations. Moreover, the statistical association between vehicle type, time of day, and violation behavior provides valuable insights for policymakers and urban traffic managers. Importantly, the analysis of rolling stop violations among road users in urban Khon Kaen reveals that vehicle type was statistically significant in its association with rolling stop violations. Stopping at crosswalks was found to have a greater correlation with rolling stop violations for motorcycles than for other vehicles. Public awareness campaigns, strict penalties for repeated offenders, and intersection redesign to improve visibility and compliance should complement AI-based enforcement systems

Practical Application: The practical implications of this study are significant for both local and national road safety strategies. In Khon Kaen, the results can inform municipal authorities and law enforcement agencies in designing targeted interventions. For example, stricter monitoring of larger vehicles and increased enforcement during evening rush hours could reduce the frequency of violations. At the policy level, the study supports the integration of AI technology into Thailand’s broader traffic management framework. Government agencies can use such systems to generate evidence-based reports, guiding legislative updates and infrastructure investments. The ability to reliably detect “rolling stop” behavior also opens new avenues for refining traffic laws, which traditionally only account for complete red-light running.

SiCal-CytoNet: A Stain-Invariant, Calibrated Multi-Scale Network for Trustworthy Cytology on SIPaKMeD

Tue, 31 Mar 2026 00:00:00 +0700

Background and Objectives: Cervical cancer screening through Pap-smear cytology remains a cornerstone of early diagnosis. However, its large-scale deployment is constrained by stain and illumination variability, morphological complexity across spatial scales, class imbalance, and unreliable confidence estimation in automated systems. Although deep learning models have achieved high classification accuracy on cytology benchmarks, many lack robustness to domain shifts and produce poorly calibrated probabilities, limiting their suitability for clinical triage. The objective of this research was to develop a stain-invariant, multi-scale, and well-calibrated deep learning framework that delivers both strong discrimination and trustworthy probability estimates for cervical cell classification, thereby enabling safe and effective decision support in screening workflows.

Methodology: This study proposes SiCal-CytoNet, a stain-aware and calibrated multi-scale network designed for Pap-smear cytology analysis. This framework employs optical-density–based stain deconvolution with controlled perturbations to address stain variability in combination with domain generalization strategies, including Fourier Domain Adaptation and MixStyle. Morphological features are learned at multiple resolutions and adaptively fused using a learnable resolution gate. A prototype-guided metric head improves class compactness under imbalance, while probability reliability is enhanced through a combination of Brier loss, evidential Dirichlet modeling, and post-hoc temperature scaling. A selective prediction mechanism based on calibrated confidence is integrated to support clinically meaningful abstention.

Main Results: On the SIPaKMeD five-class benchmark, SiCal-CytoNet achieves a test accuracy of 98.9% and a macro-F1 score of 98.7%, with excellent ranking performance reflected by an AUROC of 99.7% and an AUPRC of 99.4%. This model demonstrates strong probability calibration, attaining a low Brier score of 0.012 and an expected calibration error of 1.7%, while maintaining high discrimination under simulated stain, style, blur, and illumination shifts. Ablation experiments confirm the importance of multi-scale fusion, domain generalization, and calibration components. Using a utility-optimized selective prediction strategy, the system confidently automates 88.1% of cases while preserving high sensitivity, specificity, and predictive values for clinical triage.

Conclusions: The findings show that combining stain physics, adaptive multi-scale learning, prototype-guided classification, and explicit calibration produces a cytology model that is both highly accurate and clinically trustworthy. SiCal-CytoNet moves beyond accuracy-centric evaluation by delivering reliable probability estimates and robust performance under domain shifts, addressing critical requirements for real-world deployment in cervical cancer screening.

Practical Application: SiCal-CytoNet can be deployed as an automated triage system in cytology laboratories to assist pathologists by rapidly classifying routine Pap-smear samples while referring uncertain cases for expert review. Its calibrated outputs support transparent threshold selection, reduce screening workload, and enhance consistency across laboratories with heterogeneous staining conditions, offering tangible benefits to public health screening programs and AI-assisted diagnostic platforms.

A Recursive N-Subdivision Algorithm for Efficient Line and Curve Clipping

Bhawana Ahire, Rupali Tajanpure — Tue, 31 Mar 2026 00:00:00 +0700

Background and Objectives: Line and curve clipping is a fundamental operation in computer graphics, geometric modeling, and visualization systems, determining which portions of geometric primitives are visible within a specified clipping window. Classical clipping algorithms such as Cohen–Sutherland, Liang–Barsky, Nicholl–Lee–Nicholl, and midpoint subdivision have been widely used due to their deterministic behavior and mathematical rigor. However, these methods face increasing limitations when applied to modern graphics workloads involving large datasets, complex curves, and real-time rendering requirements. Intersection-based approaches often perform repeated and computationally expensive boundary calculations even for trivially invisible primitives, while traditional subdivision techniques may unnecessarily refine segments far from the clipping region. With the growing adoption of GPU-based pipelines, parallel processing, and curve-intensive applications such as CAD, GIS, font rendering, and scientific visualization, there is a need for a scalable, intersection-minimizing, and hardware-friendly clipping strategy. The primary objective of this research is to propose a unified and efficient clipping algorithm for both line and curve primitives that reduces redundant computations, improves scalability, and aligns with modern parallel architectures while preserving acceptable visual accuracy. RNSCA reduces redundant intersection computations, particularly for curve-heavy and large-scale datasets.Traditional algorithms may outperform RNSCA for small to moderate CPU-based line-only datasets. The primary strength of RNSCA lies in scalability and parallel suitability rather than small-scale sequential speed.

Methodology: This research introduces a Recursive N-Subdivision Clipping Algorithm (RNSCA), which generalizes classical midpoint subdivision by initially dividing each geometric primitive into N equal sub-segments. The value of N can be adaptively selected based on line length, dataset density, or available computational resources. Each sub-segment is classified using region outcodes derived from the Cohen–Sutherland framework, enabling rapid trivial acceptance or rejection through simple bitwise operations. Only ambiguous segments that potentially intersect the clipping boundary are further recursively subdivided. The recursion continues until the segment is trivially classified, a maximum recursion depth is reached, or the segment length falls below a predefined tolerance. For curve primitives such as Bézier and spline curves, parametric subdivision or polyline approximation is applied, allowing the same outcode-based classification without explicit polynomial intersection solving. The algorithm is intentionally designed to defer or avoid direct intersection calculations, relying instead on recursive refinement and classification. Experimental evaluation was conducted using synthetic datasets consisting of randomly generated line segments and curve primitives of varying sizes, and performance was compared against classical clipping algorithms using runtime, scalability, and visual accuracy metrics.

Main Results: The experimental results indicate that the proposed RNSCA effectively reduces unnecessary computational overhead, particularly for large-scale and curve-heavy datasets. For small datasets, traditional intersection-based algorithms demonstrate comparable or slightly faster performance due to lower overhead. However, as dataset size increases, RNSCA exhibits superior scalability by discarding irrelevant fragments early in the processing pipeline. In curve clipping experiments, the benefits are more pronounced, as RNSCA avoids repeated polynomial intersection evaluations and prunes a significant portion of irrelevant geometry during early subdivision stages. Across multiple test cases, the algorithm eliminated approximately 70–80% of unnecessary computations prior to deep refinement while maintaining visually consistent clipping results. Although the recursive approach may preserve very small boundary-adjacent fragments depending on the chosen subdivision parameter N, it avoids false-positive acceptance of outside geometry. The results demonstrate that RNSCA provides a flexible accuracy–performance trade-off that can be tuned for specific application requirements.

Conclusions: The Recursive N-Subdivision Clipping Algorithm represents a meaningful advancement in clipping methodologies by shifting the computational focus from intersection-centric processing to subdivision-based, outcode-driven classification. By generalizing subdivision to N segments and refining only ambiguous regions, the algorithm achieves improved scalability, reduced redundancy, and enhanced suitability for parallel processing. Its unified treatment of line and curve primitives simplifies clipping pipelines and reduces algorithmic complexity in curve-dominated workloads. While careful selection of the subdivision parameter N is required to balance precision and performance, the proposed approach complements classical algorithms rather than replacing them, offering distinct advantages in large-scale and real-time graphics environments.

Practical Application: The proposed RNSCA is particularly well suited for GPU-accelerated rendering pipelines, real-time visualization systems, CAD/CAM platforms, GIS applications, and vector graphics engines where large volumes of geometric primitives must be efficiently processed. By minimizing expensive intersection computations and enabling embarrassingly parallel execution, the algorithm reduces processing time, improves rendering throughput, and lowers computational overhead. Its adaptability allows developers and practitioners to tune precision based on application needs, making it beneficial for both interactive graphics and large-scale industrial visualization tasks. The algorithm thus offers tangible benefits to industry and research communities seeking scalable and future-ready geometric clipping solutions.

Analysis of Factors Influencing Risk Behavior of Motorcyclists: A Case Study of Chiang Mai University Students

Nima Maikanthat, Nopadon Kronprasert — Tue, 31 Mar 2026 00:00:00 +0700

Background and Objectives: Thailand continues to face high rates of road accidents, particularly among young motorcycle users who are considered as risky and vulnerable road users. The present study focused on analyzing risky riding behaviors among students at Chiang Mai University by applying the Theory of Planned Behavior (TPB) integrated with Structural Equation Modeling (SEM). The main objective was to identify key psychological and behavioral factors influencing safe motorcycle riding, specifically Attitude (ATT), Subjective Norm (SJN), Perceived Behavioral Control (PBC), Intention (INT), and Behavior (BHV) to develop an evidence-based model for improving road safety among university students.

Methodology: The study employed a combined observational and survey study approach. Quantitative data were collected from 884 students through structured questionnaires measuring TPB constructs, while quantitative data were obtained from on-campus traffic video recordings to verify the observed behaviors. SEM was used to test causal relationships among latent variables (ATT, SJN, PBC, INT, and BHV). Reliability and validity were assessed through factor analysis.

Main Results: The structural equation model exhibited an acceptable-to-good fit to the data (χ²/do = 4.93, CFI = 0.927, GFI = 0.910, TLI = 0.912, RMSEA = 0.067, RMR = 0.040). PBC had the strongest influence on both Intention (β = 0.78) and Behavior (β = 0.34), while Intention significantly mediated the relationship between PBC and safe riding behavior (β = 0.56). Attitude had a weaker impact on Intention (β = 0.16), while Subjective Norm showed no significant effect. Female and younger students (below 19) were more influenced by PBC, whereas male and older students’ behaviors were driven mainly by Intention.

Conclusions: The study concluded that PBC and Intention are the most critical determinants of safe motorcycle riding among university students. The findings suggest that improving riders’ self-efficacy, confidence, and skills is more effective than relying on social norms or peer pressure. This evidence supports using TPB as a robust framework for understanding and predicting safe riding intentions and behaviors.

Practical Application: Based on the results, a 3E-based systemic intervention—Education, Enforcement, and Engineering—under the Swiss Cheese Model is proposed to close risk gaps and strengthen road safety within the University. Practical measures include safety training programs, female rider confidence workshops, and a student reward system for safe riding behavior. These initiatives can serve as a prototype for a safe university campus.

Erratum to Navigating the Transition to Industry 5.0: Risk and Resilience in Technology Startups

Nattida Tachaboon — Tue, 31 Mar 2026 00:00:00 +0700

วารสาร Science and Engineering Connect ปีที่ 48 ฉบับที่ 4 ตุลาคม-ธันวาคม 2568 หน้า 346-369

การแก้ไขบทความ : Navigating the Transition to Industry 5.0: Risk and Resilience in Technology Startups

Erratum: Navigating the Transition to Industry 5.0: Risk and Resilience in Technology Startups

Nattida Tachaboon*
Faculty of Engineering, Chiang Mai University, Chiang Mai, Thailand
* Corresponding author E-mail: nattida.t@cmu.ac.th

URL: https://ph04.tci-thaijo.org/index.php/SEC/article/view/11667

ข้อผิดพลาดที่พบ :
ผู้เขียนขออภัยสำหรับข้อผิดพลาดในการอ้างอิงแหล่งที่มาในบทความ หน้า 347 ในส่วน Introduction บรรทัดที่ 7
เดิม
“As highlighted by Piyatamrong and Guile [2], Thailand’s innovation system suffers from structural weaknesses: most firms remain dependent on imported technologies, while government innovation policies often fail to effectively stimulate technological upgrading.”
ขอแก้ไขเป็น
“As highlighted by Intarakumnerd [2], Thailand’s innovation system suffers from structural weaknesses: most firms remain dependent on imported technologies, while government innovation policies often fail to effectively stimulate technological upgrading.”