Real Time Multi-Class Weapon and Physical Aggression Event Detection Using YOLO26 in Surveillance Videos

The rise in violence and homicide rates has been a challenge in the modern world because of the uncontrolled access to weaponry in many countries, leading to serious threats to public safety and security in general, with other consequences like psychological and economic losses. It's known that guns and other weapons are the most common form of crime incidence and violent acts [1, 2]. And as it's known, the traditional monitoring process is not very effective, especially when many monitor screens require concentration from the operator, where the personnel in charge can lose focus after short periods, leading to the possibility of missing critical events [3, 4]. All the mentioned points lead to the conclusion that human monitoring alone is not effective or accurate in the threat/danger detection process, which emphasizes the vital need for effective solutions and measures.

The recent advancement in technology and artificial intelligence (AI) specifically deep learning and computer vision, has enabled the idea of weapon detection [5]. The conventional neural networks (CNNs) proved to be very effective when it comes to object detection tasks, and when speaking about strong models in object detection many other modern frameworks stand out, like the YOLO family which gained its reputation due to its effectiveness and ability in real time with the needed balance between accuracy and speed [6, 7]. The detection task in deep learning models is highly related to many factors such as object size, lighting conditions, or occlusion, making it very difficult, especially if the images captured have complex backgrounds [8, 9]. In addition, the performance of a deep learning model depends on the dataset itself and its quality. The existing datasets are limited in size and diversity and are not accurate in representing the survival scenarios, making the model less generalizable [10, 11].

To address the limitations above, recent studies have focused on both model architecture and dataset characteristics. Newer YOLO based models enhanced the original architectures in the early versions, leading to improvement in the detection process and real-time capability [7, 8].

The current study deals with the latest progress in object detection through YOLO versions. The study contributes to enhancing the public security field in the following ways:

1. Comparing the new YOLO26-M (Medium) against YOLOv11-M, training both versions on a dataset from a previous study.
2. Constructing a dataset of 10 classes from different weapons categories and violence, containing 50k images (along with the annotations) in total.
3. Building a model from training the same version (YOLO26) on the proposed dataset capable of detecting and classifying potential threats in real time.

The sections discussed in this study are structured as follows: the literature view is summarized in Section 2, information about YOLO26 is provided in Section 3, and the methodology, containing the datasets related descriptions mentioned, the model, the environment used, and the training steps all presented in Section 4. While Section 5 discusses the results for both datasets using appropriate evaluation and consideration taken for further deployment.

Transfer learning works by letting the model reuse previously learned features; thus, it has significant performance in tasks related to computer vision. It can also be applied successfully in multi-weapon detection systems like audio-based firearm classification [22]. Focusing on real-world deployment, the YOLO models (you only look once) family proved worthy and efficient due to their high detection and fast speed. YOLO models utilize the architecture of single-stage detection, and the processing of the entire image is done as one forward pass, resulting in faster speeds than two-stage detection systems [7].

The first proposed YOLO frame work proposed by Joseph Redmon and colleagues in 2016 presented a new approach in object detection, unlike the traditional R-CNN and Faster R-CNN, which separated region proposal from classification. YOLO formulated detection as a single regression problem. By directly predicting bounding boxes and class probabilities in one forward pass through a convolutional neural network (CNN), it achieves competitive accuracy and real-time speed. Through the following years, many versions were presented, each with different architectures and suitable improvements for the specific task.

In this study the latest version of YOLO which is YOLO26, is selected as the core model for the intended work. The YOLO26 implementation used in this sudy was based on the Ultrealytics framework and the experimental configuration, software version, model size and training parameters along with the evaluation settings are reported in section 4.2 to support reproducibility. as its previous versions is also designed to deliver the needed performance and efficiency, with the capability of handling various computer vision tasks such as object detection, classification, segmentation [7]. Like the previous versions, this one comes with multiple variants, each of which can handle different tasks and deployment requirements, starting from the lightest to the heaviest in architecture nano (n), small (s), medium (m), large (l) and extra-large (x). Compared to earlier models, this one has some architectural improvements aiming to enhance real-time inference and edge deployment. This means simplified design and reduced computational complexity, all to enable as fast an inference time as possible on both CPUs and GPUs [23].

From a theoretical perspective, the proposed dataset2 contains substantial variations in object scale, illumination conditions, occlusion and visual similarity between categories, along with the complex backgrounds. Such characteristics require a capable detector that can learn robust multi-scale feature representations while maintaining efficient localization. YOLO26 is particularly suitable for this scenario because its architecture is designed to preserve detailed spatial information while simultaneously extracting high-level semantic features.

This combination is important for distinguishing visually similar categories, such as a knife and a cutter. As well as for detecting small or partially occluded objects encountered commonly in surveillance imagery.

Details about the architecture are available in the original YOLO26 publication [7], illustrating COCO mAP (50-95) versus latency performance. The more recent the model is, the higher its associated mAP, indicating the improvements over generations as well as competitive real-time detectors (PP-YOLOE+, DAMO-YOLO, and RTMDet). On the end-to-end latency axis, YOLO26 is compared with YOLOv10 and RT-DETR variants, illustrating its advantage in overall pipeline efficiency [7].

In this work, the YOLO26-M variant is selected for training and evaluation, based on the balance it provides between detection performance and computational efficiency, especially with a large-scale dataset, which makes it the most suitable for real-time surveillance, where both speed and accuracy are crucial.

This section explains the results and outcomes After data preparation, configuring the model and training, analyzing model performance using standard detection metrics with qualitative visualizations and comparative analysis.

5.1 Evaluation metrics

These are the metrics used to assess the performance of the model. Standard evaluation metrics commonly used in object detection were employed and widely adopted in YOLO-based systems, providing a comprehensive understanding of detection accuracy and robustness, serving as feedback to the model performance [23].

Precision (P) indicates how many predicted weapons are actually correct.

Precision $=\frac{T P}{T P+F P}$

Recall (R): indicator for the ability of the model to capture all weapons and danger instances.

Recall $=\frac{T P}{T P+F N}$

F1-Score: assessment of overall detection by balancing precision and recall.

$\mathrm{F} 1=2 \frac{p \cdot r}{p+r}$

Mean average precision (mAP): how well the model detects and localizes objects across all classes.

$\mathrm{mAP}=\frac{1}{C} \sum_{C=1}^C A P c$

where, c is the number of classes and APc is the average precision for the class.

Intersection over Union (IoU): measurement of the overlap between the predicted bounding box and the ground truth box.

$I o U=\frac{\text { Area of Overlap }}{\text { Area of Union }}$

These metrics are standard in modern object detection frameworks and are essential for evaluating real-time detection systems [6].

5.2 Experimental results

The trained YOLO26-M model was evaluated on both datasets using the best-performing epoch obtained during training. The evaluation was conducted on the test split, and results are presented both quantitatively and qualitatively.

5.2.1 Results of dataset1 (dangerous items dataset)

This section will cover the results according to the metrics explained briefly for each class and the model, as shown in Table 6, along with the visualization sample, the curves and the confusion matrix. Comparison results are shown in Table 7 between omiotek and z. Model [21] and the proposed model in this study on the same dataset.

Table 6. Dataset1 classes and model results

The model achieved (0.954) overall precision, indicating a high ability to identify weapon instances correctly with minimal false positives and this is very important in surveillance systems where the false alarms must be kept as low as possible. The recall value was found to be (0.848), meaning that while the model detects most weapon instances correctly, some of the objects may still not be recognized and may be missed, aligning with the known challenges in the weapon detection field [8]. Overall mAP@50 of 0.916 proves strong capability for detection, while the mAP@50-95 of 0.773 reflects the performance under stricter localization criteria. The range between the two is a strong indicator for further improvement in bounding box precision. Among the results, we notice some high numbers for the AR and bat in terms of performance and the reason is their large sizes and special features, but in contrast, the handgun showed lower mAP, which backs up the literature indicating small objects are harder to detect accurately [8, 9].

Figure 2 shows performance curves providing further evaluation of models' rating, starting with (a) the precision-recall curve. A strong overall performance is obvious, meaning that the model keeps high precision across a wide range of recall values. The precision-confidence curve (b) shows the increase in precision with higher confidence thresholds, reflecting the reliability of high-confidence predictions. on the contrary, the recall confidence curve (c) shows that recall is highest when the confidence is at the lowest thresholds and decreases when the threshold increases. which highlights that there is a trade-off between avoiding false positives and detecting more objects. The F1 confidence curve (d) provides an optimal balance between precision and recall at a threshold of confidence of nearly 0.5, achieving the maximum F1 score of 0.90 overall. The analysis of these curves confirms that the model is stable, effective and suitable for real-time deployment. 

3_2.png

Figure 2. Performance curves for dataset1 model, (a) precision-recall curve, (b) precision-confidence curve, (c) recall-confidence and (d) F1-confidence curve

The confusion matrix is shown in Figure 3, providing per-class prediction performance and misclassifications, noticing diagonal values where the numbers are a good indicator for strong classification accuracy, starting from the highest values (bat = 0.92) and then (AR = 0.91) and the (handgun = 0.87). Achieved the highest correct prediction rates. Knife and machete showed lower accuracy (0.84 and 0.80, respectively), indicating difficulty distinguishing these two classes. This is due to the similarities between these objects: both have overlapping shapes and similar features. In addition, a portion of the objects is misclassified across all classes, indicating missed detections. Relatively, the low values of off-diagonal means indicate limited inter-class confusion, and most errors are not related to incorrect classification but to missed detections.

3_3.png

Figure 3. Confusion matrix for dataset1 model

Table 7 shows a comparison of the performance metrics between the two models (previous study v11-m) [21] and the proposed YOLO26-m on the dangerous items study dataset.

Table 7. Comparison of YOLOv11-M and YOLO26-M

Table 7 demonstrates the improvement in precision (+5.8%), considered an indicator for weapon and non-weapon object discrimination, reducing false positives. Recall decreased compared with YOLOv11, indicating that YOLO26-M produced fewer but more confident detections. This trade-off may reduce false alarms but may also increase missed detections; therefore, confidence threshold tuning is necessary, depending on whether the deployment prioritizes sensitivity or precision. More importantly, noticing the improvement in mAP @50-95 highlights the enhanced localization accuracy across multiple (IoU) thresholds, reflecting the advancement of the YOLO26 architecture, particularly the ability to balance efficiency and accuracy for real-time tasks. All these results confirm that YOLO26 is a strong candidate for real-time applications.

5.2.2 Results on dataset2 (proposed dataset)

This section is concerned with the trained YOLO26-M on the proposed dataset2, consisting of 50,000 images from multiple categories and by using the metrics defined earlier, the evaluation was done on the test split set. It should be noted that the violence class is a physical aggression event category and differs conceptually from the weapon classes, as it represents a visually observable threat-related event rather than a physical object. Nevertheless, it was included to evaluate the capability of the proposed framework to detect multiple forms of security-relevant threats within a unified surveillance setting. Table 8 shows the exact results.

Table 8. Results for dataset2

The model achieved an overall precision of 0.954 and recall of 0.913, resulting in an F1-score of 0.933, which indicates a strong balance between detection accuracy and completeness. In addition, both mAP values demonstrate robust performance across both thresholds.

In Figure 4, the normalized confusion matrix is considered a strong performance for the classification operation, where some true positive rates are noticed in some classes like cutter axe AR and shotgun; other classes like hammer violence and knife also show reliable detection and classes like scissor and rocket launcher have slightly lower accuracy, suggesting that these classes are challenging to distinguish. Misclassification is minimal between the classes, indicating the model learns special features for each category.

3_4.png

Figure 4. Confusion matrix for dataset2 model

However, some of the samples is misclassified as background (up to 0.18 for scissors and 0.13 for rocket launcher). Overall, the matrix confirms that the model achieves high classification accuracy with limited confusion for inter-class. And most errors are associated with missed detections rather than incorrect class predictions.

Figure 5 shows evaluation curves where similar trends were observed in dataset1 with improved performance. The Precision–Recall curve (a) shows strong overall detection, with a high mAP@0.5 of 0.955, indicating that the model maintains high precision across most levels of recall. The Precision–Confidence curve (b) reveals that precision improves steadily along with confidence confirming the reliability of high-confidence predictions. In contrast, the recall–confidence curve (c) shows that recall is highest at lower confidence levels (nearly 0.95) and decreases as the threshold increases. The F1–Confidence curve (d) identifies that the model achieved higher precision across confidence thresholds and a higher F1 score at a lower optimal threshold (0.335). Overall, the curves tell how model performance is considered robust and suitable for real-world applications.

3_5.png

Figure 5. Performance curves for dataset2 model, (a) precision-recall curve, (b) precision-confidence curve, (c) recall-confidence and (d) F1-confidence curve

5.2.3 Cross-dataset validation using YouTube Gun Detection Dataset

Although dataset1 already provides direct comparison against previously reported YOLO family versions from the original study, an additional evaluation was conducted on the independent YouTube Gun Detection Dataset (GDD) benchmark dataset to investigate the generalization behavior of recent YOLO generations under different surveillance scenarios. This supplementary experiment included YOLOv8, YOLOv11 and YOLO26 trained and evaluated under identical settings using the official dataset protocol; the results are summarized in Table 9.

The YOUTUBE GDD benchmark provides complementary evidence rather than a direct claim of universal superiority. YOLO26 achieved the highest overall precision and the best mAP@50-95, indicating improved localization under stricter IoU thresholds. Therefore, the GDD results indicate that YOLO26 provides a competitive accuracy-efficiency trade-off, especially in localization robustness. This balance supports the use of YOLO26 as a practical detector.

Table 9. Comparative benchmark evaluation on Gun Detection Dataset (GDD)

5.2.4 Failure cases and limitations analysis 

Although a strong overall performance is achieved by the proposed framework across datasets, several challenges were identified during the qualitative inspection of the prediction results. Most of the failure cases are associated with small object size, partial occlusion and complex background. In addition to the low illumination conditions and visual similarities between classes, as illustrated in Figure 6. 

3_6.png

Figure 6. Failure cases samples

For dataset1, some confusion was observed between the knife and the machete due to their similar geometric characteristics and features. In addition, handgun detection occasionally failed when the object occupied only a small portion of the image, which is consistent with findings reported in previous weapon-detection studies.

And for dataset2, the scissor and rocket launcher classes produced lower recall compared to the remaining classes. Scissors often appeared as thin objects with limited information, making them more difficult to localize accurately. Rocket launchers presented an additional challenge due to their diverse appearance, viewing angles and lower representation in publicly available surveillance imagery.

The violence category also exhibited greater variability than conventional weapon classes because the detection here depends on human posture and interaction patterns rather than a clearly defined physical object, leading to some misdetections occurring in the complex scenes containing multiple individuals.

These observations indicate that future improvements may be achieved through additional training samples representing small-scale objects, enhanced processing for low-light images and the incorporation of temporal information for behavior-related event detection.

5.3 Speed analysis and real-world deployment 

An NVIDIA H100 GPU (80GB HBM3) was used to infer the performance of the proposed model of YOLO26-M using 960x960 pixels of input resolution, which is consistent with the training configuration. The two datasets are evaluated and considered for preprocessing, inference, and post-processing stages.

For dataset1, the measured latency was approximately 2.7 ms per image using the NVIDIA H100 environment. These results demonstrate highly efficient inference performance under high-performance computing conditions while maintaining strong accuracy for detection. These findings are considered an indicator of the suitability of the proposed framework for latency-sensitive objects, although the deployment performance is expected to vary depending on the hardware platform in use.

For dataset2, high efficiency execution was demonstrated; the time measured as follows: (Preprocessing: ~0.1 ms, Inference: ~1.7 ms, Postprocessing: ~0.1 ms). This performance indicates that the model operates well beyond conventional real-time requirements.

The experimental evaluation was conducted on high-performance data-centre hardware (NVIDIA H100), representing an upper bound of achievable performance. To take the assessment of deployment feasibility beyond the data-center environment, the trained YOLO26 model on dataset2 was additionally evaluated on consumer-grade hardware (NVIDIA GeForce RTX 3050 Laptop GPU, 6 GB VRAM) using the same split and input resolution (960 × 960). The model maintained comparable detection performance, achieving a precision of 0.953, a recall of 0.913, a mAP@50 of 0.954, and a mAP@50-95 of 0.840, as shown in Table 10. These results indicate that the proposed framework preserves its detection capability on lower-resource hardware, although inference throughput is naturally reduced compared with the H100 platform. In practical deployment scenarios, the system is expected to operate on more resource-constrained platforms such as NVIDIA RTX-series GPUs and embedded systems (e.g., Jetson family). On such platforms, an increase in inference latency is expected due to reduced computational resources. However, the achievable inference speed remains dependent on computational resources that are available, deployment configuration and optimization strategy adopted for the target platform.

Table 10. Hardware results comparison

The deployment benchmark demonstrates that the proposed framework preserves its detection performance across different hardware platforms. While inference latency increased from approximately 1.9 ms on the NVIDIA H100 to 98.6 ms on the RTX 3050, the rest of the values remained unchanged, indicating that the reported H100 measurements represent an upper-bound performance estimation, where the RTX provides a more realistic deployment scenario for simple consumer-grade hardware. 

Figure 7 illustrates a practical deployment scenario for the proposed surveillance framework, in which the YOLO26-M detector processes the inputs captured by the cameras. Predictions exceeding a predefined confidence threshold are treated as potential threats and generate alerts for further inspection. and to reduce the impact of false detections, human verification is recommended before initiating any security-related actions. Once the threat or danger is confirmed, the system can assist security personnel by providing rapid notification and situational awareness for appropriate response measures. 

3_7.png

Figure 7. Proposed deployment workflow

For practical deployment, additional optimization techniques can be applied, like Quantization (FP16 / INT8) or TensorRT acceleration and even Batch size tuning. These techniques are widely used to reduce latency and improve throughput without significant degradation in detection performance.

From an architectural perspective, the use of high input resolution (960 × 960) improves detection of small objects and partially occluded weapons. However, it also increases computational cost. Therefore, deployment configurations could be taken with considerations such as reducing input resolution and/or selecting lighter model variants (e.g., YOLO26-S or YOLO26-N) to achieve a better balance between speed and accuracy depending on hardware constraints.

5.4 Ethics and privacy considerations

The development of intelligent systems requires careful considerations related to ethical, privacy and security implications. The proposed framework is designed to improve public safety through early detection, but to prevent misuse and protect individual rights, its deployment must be accomplished with appropriate safeguards. 

All the datasets used in this study were collected from public sources, existing research, online repositories, and manually acquired images intended exclusively for research and academic purposes. Most of the dataset information focuses on weaponry from different angles to make the recognized personnel less identifiable, and personally identifiable information was not intentionally collected and the focus of the annotation process was restricted to threat-related objects and activities. In practical deployment scenarios, some additional privacy-preserving measures should be considered, such as video anonymization and face blurring to minimize unnecessary exposure of personal information. 

Furthermore, the proposed system should be viewed as a decision-support tool rather than a fully autonomous security solution. The outcomes of the detection process may be affected by environmental conditions, object occlusion, lighting variations, or even visual similarities between similar classes. Therefore, human-in-the-loop is heavily recommended before initiating security-related actions or interventions. This approach can reduce the impact of false alarms and improve operational reliability.

Also, it should be noted that access to surveillance data and detection results should be controlled through appropriate cybersecurity and data protection mechanisms, preventing unauthorized access or misuse. Future deployments should also comply with applicable privacy regulations, institutional policies, and ethical guidelines governing the use of AI in public environments.

By taking these considerations seriously and incorporating them, intelligent weapons detection systems can contribute to public safety while maintaining transparency, accountability and respect for individual privacy.