Object recognition and tracking have long been a challenge, drawing considerable attention from analysts and researchers, particularly in the realm of sports, where it plays a pivotal role in refining trajectory analysis. This study introduces a different approach, advancing the detection and tracking of soccer balls through the implementation of a semi-supervised network. Leveraging the YOLOv7 convolutional neural network, and incorporating the focal loss function, the proposed framework achieves a remarkable 95% accuracy in ball detection. This strategy outperforms previous methodologies researched in the bibliography. The integration of focal loss brings a distinctive edge to the model, improving the detection challenge for soccer balls on different fields. This pivotal modification, in tandem with the utilization of the YOLOv7 architecture, results in a marked improvement in accuracy. Following the attainment of this result, the implementation of DeepSORT enriches the study by enabling precise trajectory tracking. In the comparative analysis between versions, the efficacy of this approach is underscored, demonstrating its superiority over conventional methods with default loss function. In the Materials and Methods section, a meticulously curated dataset of soccer balls is assembled. Combining images sourced from freely available digital media with additional images from training sessions and amateur matches taken by ourselves, the dataset contains a total of 6331 images. This diverse dataset enables comprehensive testing, providing a solid foundation for evaluating the model’s performance under varying conditions, which is divided by 5731 images for supervised system and the last 600 images for semi-supervised. The results are striking, with an accuracy increase to 95% with the focal loss function. The visual representations of real-world scenarios underscore the model’s proficiency in both detection and classification tasks, further affirming its effectiveness, the impact, and the innovative approach. In the discussion, the hardware specifications employed are also touched on, any encountered errors are highlighted, and promising avenues for future research are outlined.

Given that approximately 70% of traffic accidents are attributable to driver-related factors, it is necessary for vehicles to incorporate technologies that reduce risk through preventive actions derived from traffic-scene analysis. Interpreting the driving environment is non-trivial and is commonly decomposed into sub-tasks; among them, traffic light perception is critical due to its role in regulating vehicular flow. This paper evaluates five YOLO CNN families (YOLOv8–YOLOv12) on two tasks: (i) traffic light detection and (ii) traffic light state recognition (green, yellow, red). The evaluation uses a hybrid dataset comprising the public LISA traffic light dataset and a custom dataset with images from Mexico City captured under diverse lighting conditions—a relevant setting given the city’s high traffic intensity. The results show mAP@0.50 = 94.4–96.3% for traffic light detection and mAP@0.50 = 99.3–99.4% for traffic light state recognition, indicating that modern YOLO variants provide highly reliable performance for both tasks under natural illumination variability.

We present a modular multi-attribute vehicle analysis pipeline that integrates YOLO-based models and an OCR engine into a single workflow. The system detects vehicles, classifies color, recognizes make and sub-model, detects license plates, and extracts plate characters to generate a structured vehicle record. Vehicle detection is reported with standard metrics (precision, recall, and mAP@0.5), while license plate detection is reported at IoU = 0.3 to reflect the small-object nature of plates and downstream OCR usability. Among the evaluated versions, YOLOv8 provides the most balanced overall performance across modules, while maintaining real-time-equivalent throughput of approximately 18–22 FPS for the full pipeline on recorded traffic videos, depending on scene complexity. We emphasize module-level evaluation and runtime benchmarking; instance-level end-to-end identification across unique vehicles is defined as future work once track-based ground truth becomes available.

Urban road user detection supports safety-critical applications in Intelligent Transportation Systems, but architecture assessment requires balancing predictive performance with computational efficiency under controlled and leakage-aware evaluation conditions. This study adopts a three-stage benchmarking framework for vehicle, motorcycle, and pedestrian detection using a hybrid dataset of 4983 images derived from LISA and traffic scenes captured at 18 complex intersections in Mexico City. Stage I provides an exploratory comparison of six medium-sized YOLO architectures (YOLOv8m, YOLOv9m, YOLOv10m, YOLOv11m, YOLOv12m, and YOLOv26m), whereas Stage II restricts leakage-controlled confirmatory evaluation to three selected architectures: YOLOv10m, YOLOv11m, and YOLOv26m. Stage III examines the corresponding YOLOv10n and YOLOv26n nano variants. Both sources were manually annotated under a unified three-class protocol, producing 3483 quantitative images with 6624 ground truth instances and 1500 images for qualitative inference. Annotation reproducibility reached a mean pairwise IoU of 0.85 ± 0.11 and Fleiss’ κ of 0.968. A retrospective audit of the original image-level split identified acquisition group overlap and confirmed near duplicates across partitions. A corrected group-disjoint split eliminated shared groups, exact duplicates, and confirmed near duplicates while preserving the aggregate source and class distributions. Under three-seed leakage-controlled confirmatory evaluation, YOLOv26m achieved the highest mAP@0.50:0.95 among the three Stage II confirmatory models (0.5717 ± 0.0013), followed by YOLOv11m (0.5632 ± 0.0012) and YOLOv10m (0.5516 ± 0.0014), with paired bootstrap analysis supporting all three pairwise differences after Holm correction. Source-specific evaluation revealed consistently lower performance on the Mexico City subset than on LISA. YOLOv11m showed the lowest overall calibration error (ECE = 0.052 ± 0.003). At the nano-scale, YOLOv10n provided higher mAP@0.50:0.95, lower latency, and greater throughput than YOLOv26n, whereas YOLOv26n required fewer GFLOPs and less benchmark energy. The Stage II group-disjoint protocol produced more conservative performance estimates than Stage I; however, the observed mAP reduction cannot be quantified as the effect of leakage removal alone because the two stages also differed in optimization configuration and repeated-run design. These findings support multi-objective model selection based on predictive performance, calibration, latency, complexity, memory, energy consumption, and deployment constraints.

Background/Objectives: Lung cancer outcomes depend strongly on early detection, highlighting the need for accurate, robust, and interpretable methods for identifying pulmonary nodules on computed tomography (CT). This study proposes a multi-organ anatomical-contextual framework based on YOLOv8–YOLOv12, in which supporting thoracic structures are incorporated as spatial references during training while the pulmonary nodule is retained as a single invariant pathological class. Methods: The models were trained and internally validated on 5000 2D axial CT slices from 320 patients in the LUNG-PET-CT-DX dataset and externally evaluated on an independent LIDC-IDRI subset comprising 700 axial slices from 70 patients, without retraining or fine-tuning. A controlled ablation study comparing Nodule-only with Nodule + Anatomical Context was performed. Results: The anatomical-contextual scheme showed consistent improvements across all five architectures, with mean absolute gains of +0.078 in Precision, +0.122 in Recall, +0.101 in F1-score, +0.079 in mAP@50, and +0.190 in mAP@50–95. During internal validation, YOLOv10m achieved the highest Precision (0.979) and mAP@50 (0.987), as well as the shortest inference time (9.25 ms), whereas YOLOv11m achieved the highest Recall (0.971) and F1-score (0.962). During external evaluation, YOLOv10m retained the highest Precision (0.924) and mAP@50 (0.956), while YOLOv8m achieved the highest Recall (0.930) and mAP@50–95 (0.845); both models achieved an F1-score of 0.922. Conclusions: These findings provide experimental evidence that anatomical-contextual learning improves pulmonary nodule detection while maintaining competitive performance across datasets, supporting its potential as an anatomically informed and interpretable strategy for AI-assisted thoracic CT analysis.