1. Introduction
Amid growing societal concerns regarding the accelerating effects of climate change and the transition toward a sustainable economy, global interest in electric vehicles (EVs) has substantially increased. EVs do not produce direct emissions of environmental pollutants such as carbon dioxide (CO₂), nitrogen oxides (NOx), and particulate matter (PM), which are typically produced by internal combustion engines. Given their reliance on electricity as a primary energy source, these vehicles have attracted attention as an environmentally friendly mode of transportation, and their adoption is expected to reduce carbon emissions while concurrently lowering fuel expenditures [1,2].
Reflecting this trend, the global automotive industry is increasing production of EVs. This fundamental shift is reshaping automotive markets, which heretofore have always been based on vehicles with engines. In 2024, the market size served by the global automotive industry was valued at approximately USD 4.35998 trillion and is projected to grow to approximately USD 6.67828 trillion by 2032. This figure signifies that the compound annual growth rate (CAGR) is predicted to exhibit an average value of approximately 5.66% from 2024 to 2032. These statistics underscore the pivotal role of shifts in demand centered on eco-friendly vehicles in catalyzing market expansion According to statistics from the International Energy Agency (IEA), EVs represented approximately 18% of global sales of new cars in 2023. The national transition policies and declared targets indicate an increase to approximately 35% by 2030. This suggests a substantial expansion of the EV market share over the next decade. The Ministry of Science and Information and Communication Technology reports that the South Korean government aims to achieve a 10% global market share by 2030 via the production of 3 million EVs and 850,000 hydrogen vehicles [3]. The increased prevalence of EVs has led to a concomitant rise in the usage of lithium-ion batteries as a crucial component in EV systems. Batteries containing rare metals such as lithium, nickel, and cobalt retain partial performance even after the completion of their service life. Therefore, recycled EV batteries possess high economical value because they can be repurposed as energy storage systems (ESS) after the end of their ordinary life cycle [4]. Consequently, the development of methods for EV battery reuse and resource recovery has become a key focus of research to improve resource circulation and energy efficiency.
The accelerated adoption of EVs has precipitated a substantial surge in demand for lithium-ion batteries, raising concerns regarding their safety in road vehicles [5]. Lithium-ion batteries can provide high performance owing to their high energy density; however, thermal runaway events triggered by external impacts, overcharging, internal short circuits, and manufacturing defects remain a significant concern. These events can result in fire incidents, which are often accompanied by chain explosions. These fires pose a significant safety hazard because they engender more extensive damage than conventional fire events involving vehicles with internal combustion engines and present greater challenges during suppression. Specifically, EV battery fires exhibit rapid spreading once ignited as well as a high potential for reignition. These characteristics substantially increase the difficulty of effectively controlling such fires employing existing fire suppression equipment and response systems.
The recent, recurring occurrence of EV fire accidents accidents has the potential to undermine consumer confidence in EV safety and slow adoption rates [6,7]. These fires pose a significant safety hazard because they engender more extensive damage than conventional fire events involving vehicles with internal combustion engines and present greater challenges during suppression [5]. Consequently, there is a pressing need to develop technological solutions that can detect EV fires at an early stage and alert first responders to the danger. At present, new research is focusing on the development of AI-based fire detection systems that can perform real-time video analysis. In particular, the You Only Look Once (YOLO) algorithm has demonstrated notable efficacy in real-time object detection from image and video data owing to its high computational efficiency and fast processing speed. This algorithm has been applied in several fire detection scenarios.
This study compared and analyzed the performance of EV fire detection models by applying bounding-box and segmentation labeling techniques to the n, s, m, l, and x models of YOLOv11 to address limitations such as detection latency, low accuracy, and sensitivity to environmental variability. This approach enabled a comprehensive evaluation of their performance across models with diverse architectures. Based on this process of iterative development, we propose a model specifically optimized for EV fire detection.
2. Overview of the YOLOv11 Algorithm
2.1. YOLOv11 architecture
The graph shown in Figure 1 compares the performance and inference latency of various object detection models. Recent studies have increasingly focussed on lightweight deep learning models that simultaneously achieve high accuracy and high inference speeds in object detection based on the COCO dataset. The YOLOv11 model was developed to address this demand. Building on the existing YOLOv8 architecture, YOLOv11 improves computational efficiency and detection accuracy by introducing cross convolution with 2 × 2 kernels (C3k2) and cross-stage partial with spatial attention (C2PSA) modules. YOLOv11 demonstrated superior performance in terms of mean average precision (mAP) compared with the existing YOLOv8 model on the COCO dataset. This enhancement was achieved in conjunction with a reduction approximately 22% in the number of parameters in the entire architecture. Additionally, the inference speed of YOLOv11 was reported to have improved by about 15-20% [8,9,10]. The network is structured identically to the YOLO series, comprising backbone, neck, and head modules. At each stage of the process, architectural improvements are applied to enhance the representation of multi-scale features and detection performance for small objects. Figure 2 presents the architecture of the YOLOv11 model.
The backbone of the network extracts feature maps at multiple spatial resolutions from the input image, and the C3k2 architecture has been introduced to replace the C2f module used in YOLOv8. C3k2 is a lightweight convolution structure based on a 2 × 2 kernel, which is designed to reduce computational load while effectively preserving spatial details. This structure is based on the cross-stage partial (CSP) architecture, which uses only a portion of the input feature map for computations to enhance computational efficiency, minimize information loss, and improve detection performance for small objects [8,9].
The neck serves as an intermediate processing stage that fuses feature maps of multiple resolutions extracted from the backbone before forwarding them to the head. Both the C3k2 and C2PSA modules are applied during this process. C2PSA is an extension of the existing partial spatial attention (PSA) architecture designed to detect objects in images containing complex backgrounds and objects that are occluded by separating spatial attention and channel interaction into independent paths for parallel processing. Furthermore, the upsampled feature map is seamlessly integrated with high-resolution information to detect objects with diverse dimensions [8,9].
The head of the model predicts object positions, objectness scores, and class probabilities. This component comprises a multi-stage C3k2 block and a detection layer. The C3k2 block enhances detection sensitivity for small or densely packed objects by integrating depth-specific feature information, whereas the detection layer outputs the coordinates of the obtained bounding box in conjunction with the classification results. The head incorporates a convolution-batch normalization-sigmoid linear unit (SiLU) activation function (CBS) module to reliably perform feature map normalization and nonlinear transformation while enhancing the convergence speed of the network and the stability of the obtained predictions. During the post-processing stage, the non-maximum suppression (NMS) technique is applied to eliminate redundant predictions and enhance the reliability of the final results [8,9].
2.2. Bounding boxes and segmentation
The process of labeling plays a critical role in the detection and distinction of objects within model images. The labeling methods employed in this context include both bounding-box and segmentation labeling. Figure 3(a) shows the original image, while (b) shows the bounding-box labeling method, and (c) depicts the segmentation labeling method.
Bounding-box labeling defines the smallest rectangular area that contains an object to indicate object position and size. Its simple structure and fast labeling speed enable efficient computational processing. Therefore, this approach is the standard choice across several deep learning-based object detection frameworks, including YOLO and R-CNN variants. Nevertheless, background elements other than objects can be included within the bounding box, which can introduce unnecessary information that may influence model training [9].
By contrast, segmentation labeling delineates the position and shape of objects on a pixel-by-pixel basis to facilitate a more precise representation of their form. This method enables a distinct separation between objects and backgrounds to reduce the possibility of detection errors caused by background interference. Segmentation labeling has demonstrated the capacity to maintain high precision and accuracy even with complex input images featuring overlapping objects or environments characterized by a high density of visually similar objects [9].
Tables 1, 2 present the performance metrics of YOLOv11 using bounding-box and segmentation labeling, respectively. The YOLOv11 model utilized in this study offers five variants of different sizes, designated as n, s, m, l, and x, which are derived via compound scaling of network depth and width while preserving the integrity of the network architecture. The hierarchical depth of the representation capacity and the effective receptive field generally increases in order of n, s, m, l, and x models, resulting in progressively improved accuracy and robustness. However, this augmentation is accompanied by a proportional increase in memory utilization as well as the duration of training and inference processes. Employing bounding box labeling, the YOLOv11n model achieved mAP (box) of 78.4 with 2.7 million parameters and 17.2 billion floating-point operations (FLOPs). The most complex model, YOLOv11x, achieved an mAP of 81.3 with 58.8 million parameters and 520.2 billion FLOPs. Under segmentation labeling, YOLOv11n attained an mAP (box) of 38.9 and an mAP (mask) of 32.0, while exhibiting an inference speed of 1.8 ms in the TensorRT environment. The YOLOv11x model exhibited an mAP (box) of 54.7 and an mAP (mask) of 43.8. These findings indicate that both methods demonstrated enhanced accuracy with increasingly large models, albeit at the cost of higher computational loads and longer processing times. Therefore, careful selection of the model scale and labeling method based on application-specific requirements remains essential for practical deployment [10,11,12,13].
3. Experiment Details
3.1. Experimental setup and environment
A total of 3,000 images depicting EV fires were collected from the Roboflow platform as well as from domestic and international news articles. The collected data were used to train the EV fire detection model. Figure 4 presents representative examples from the collected dataset. The images were labelled using the Polygon Tool and the Smart Polygon (AI Labeling) features of the Roboflow platform. The Polygon Tool facilitates manual delineation of an object's outline for high-precision labeling; however, significant time is required to complete the labeling process. Conversely, the Smart Polygon tool is based on the Segment Anything Model (SAM); it spontaneously identifies the expected boundaries of an object selected by the user to reduce the time required to label images. Moreover, the initially generated mask can be subject to further refinement through additional editing.
Following labeling, all images were resized to a resolution of 640 × 640 pixels and preprocessed to enhance computational efficiency during the training and inference processes while preserving the necessary details for object detection. To evaluate the stability of the training and the generalization performance of the model, the full dataset was split into three subsets including a training dataset comprising 2,400 images (80%), validation dataset containing 300 images (10%), and testing dataset consisting of 300 images (10%).
As shown in Table 3, we utilized a specific configuration of hyperparameters and a particular hardware environment to the train the object detection model. All training and validation were performed in a CUDA 12.3 environment using PyTorch version 3.9.0. The hardware utilized for the training comprised a high-performance computing environment based on a system with an Intel® Xeon® Silver 4210 10-core central processing unit (CPU), 192 gigabytes of random access memory (RAM), and two Nvidia RTX 3090 graphics processing units (GPUs). All hyperparameters were set identically for a meaningful comparison. Stochastic gradient descent (SGD) optimization was applied according to the default configuration of YOLOv11 with a fixed learning rate of 0.01. The number of epochs was set to 300, and the batch size was set to 16 to optimize GPU utilization. The resolution of the input images was normalized to 640 × 640 pixels across all models to minimize differences in performance owing to variations in the input conditions.
3.2. Experimental results
Two object detection labeling methods were implemented using the n, s, m, l, x models of YOLOv11. Subsequently, their detection performance for images of EV fires and smoke were quantitatively evaluated and compared using both standard performance metrics and qualitative visualizations. The performance of the models was evaluated by calculating the intersection over union (IoU) threshold (τ). In this context, precision is the rate of accurate detection, recall is the detection rate among actual objects, and the F1-score is the harmonic mean of these two measures. The mAP is the mean average precision across classes, whereas the mAP50 is the mAP at τ = 0.50 and mAP50-95 is similarly the average mAP between τ = 0.50 and τ = 0.95. The speed at which the network could output inferences was measured in terms of frames per second (FPS). Tables 4, 5 respectively present performance results for the bounding-box and segmentation labeling methods of the YOLOv11-based object detection model.
The experimental results of the bounding-box labeling technique shown in Table 4 demonstrate that the YOLOv11n model achieved high accuracy with an mAP50 of 0.7736 and mAP50-95 of 0.5017 despite its lightweight architecture. Furthermore, it exhibited the highest inference speed with a frame rate of 44.6167 FPS. The YOLOv11s model attained an FPS of 42.5812 and an F1-score of 0.7622, which indicates an optimal balance between precision and speed. The YOLOv11m model demonstrated overall stable performance with an mAP50 of 0.7663, mAP50-95 of 0.4900, and F1-score of 0.7601. Among these models, the YOLOv11l model exhibited the most balanced performance between precision and recall with a precision of 0.8448, a recall of 0.7657, and an F1-score of 0.7577. This finding indicates that the model is suitable for environments that require high accuracy. By contrast, the YOLOv11x model exhibited the highest number of parameters, with an mAP50 of 0.7672 and mAP50-95 of 0.4955. However, the FPS decreased to 31.5061, which indicates that an increase in the complexity of the model does not necessarily guarantee enhanced performance. Therefore, selecting a model that effectively balances computational efficiency and accuracy is crucial.
The detection results presented in Table 5 demonstrate that the mAP values were marginally lower overall than those obtained using the bounding-box labeling method. This discrepancy can be attributed to the disparity in evaluation criteria, which is a direct consequence of the intrinsic capacity of the labeling method to accurately delineate boundaries and shapes at the pixel level. However, segmentation labeling offers distinct advantages in specific applications because it can reflect the shape and position of actual objects more accurately.
The YOLOv11s model exhibited the highest inference speed, with an mAP50 of 0.7613, mAP50-95 of 0.4375, and F1-score of 0.7557, and inference speed of 40.5783 FPS, which substantiates the applicability of this model for real-time environments. The YOLOv11l model exhibited notably high performance with a precision of 0.8514, a recall rate of 0.7143, and an F1-score of 0.7616. This performance indicates the suitability of this model for high-precision detection tasks. The YOLOv11n and YOLOv11m models exhibited consistent performance, with mAP50 values of 0.7652 and 0.7593, respectively. Additionally, both models demonstrated mAP50-95 values of 0.4445 and 0.4444, in conjunction with F1-score values of 0.7556 and 0.7510, respectively. By contrast, the YOLOv11x model demonstrated the most optimal performance, with a recall of 0.7339, an mAP50 of 0.7667, and an mAP50-95 of 0.4449. However, its frame rate was 27.8733 FPS, which was the lowest among all of the models.
Figure 5 shows the visualization results for the bounding-box labeling, and Figure 6 shows the visualization results for segmentation. The visualization results presented in Figures 5, 6 demonstrate patterns consistent with the quantitative metrics. The bounding-box labeling method enabled rapid detection of fire and smoke; however, incomplete coverage frequently occurred for objects with ambiguous boundaries or high diffusion such as smoke. In scenes where smoke dispersed beyond the confines of the vehicle, the area detected by the bounding box was narrower than the actual smoke.
By contrast, the implementation of the segmentation labeling technique facilitated the delineation of object boundaries and diffusion ranges within a single scene to enable object recognition that more closely resembled real conditions. For images depicting flames spreading along building exteriors, segmentation accurately captured both the extent of fire propagation and the spatial distribution of smoke, which remained clearly distinguishable from the background. This finding suggests that the segmentation labeling method is a viable approach to detect objects in challenging situations such as complex environments or overlapping objects.
4. Conclusion
In this study, we implemented bounding-box and segmentation labeling using five variants (n, s, m, l, and x) of the YOLOv11 object detection model. A comparative performance analysis was conducted to evaluate the performance of the models in detecting EV battery fires and smoke. Experiments were conducted under identical learning environments and hyperparameter settings; precision, recall, mAP50, mAP50-95, F1-score, and frame rate in FPS were recorded as key measures of performance. We also examined the accuracy and computational efficiency of the different models.
The bounding-box labeling method generally exhibited superior performance in terms of inference speed compared with the segmentation labeling method. In particular, the YOLOv11n and YOLOv11s models processed the input at high speeds while retaining high precision, attaining FPS values of 44.62 and 42.58, respectively, in conjunction with mAP50 values of 0.7735 and 0.7693. The YOLOv11l model exhibited the highest precision, with a recorded precision value of 0.8448 and an F1-score of 0.7577. This performance indicates that the model is suitable for environments where precise detection is required. Despite possessing the most parameters and a complex architecture, the YOLOv11x model achieved an mAP50 of 0.7672, which was 0.63% lower than YOLOv11n. It also operated at 31.51 FPS, approximately 29% slower, which indicates lower computational efficiency.
With respect to the segmentation labeling methods, the mAP50 for the bounding-box labeling method based on YOLOv11n was 0.7735, whereas that for the method based on segmentation was 0.7652, which is 0.83% lower. The mAP50-95 values were 0.5016 and 0.4445, respectively, which represents a decrease of approximately 5.7%. This discrepancy is attributed to the precise labeling structure based on pixel-level masking. Quantitative metrics revealed that the overall difference in performance between the two methods was generally within the range of 1 to 6%, which indicates that the difference was not significant. However, an analysis of the visualization results indicated notable limitations of the bounding-box labeling method. Specifically, rectangular representations inherently include background regions surrounding smoke, which can resulted in false positives or negatives. Conversely, segmentation labeling precisely delineates the boundaries of smoke and initial flames for each pixel of the image. Therefore, it minimizes superfluous background interference and alleviates the variation in confidence between frames. Furthermore, the model was able to identify states that exceeded an established threshold, which suggests that it has some promise in fire detection based on observed shapes. These results can be interpreted as indicating the validity of the model in real-world environments, which is challenging to assess using solely simple quantitative metrics. Therefore, the segmentation labeling method is considered appropriate for application environments where rapid and precise initial detection is paramount such as EV battery fires. In future research endeavors, we plan to employ a range of generative artificial intelligence systems to generate diverse datasets comprising fire images. Subsequent studies will be conducted to enhance the generalization performance of various artificial intelligence models.
PDF Links
PubReader
ePub Link
Full text via DOI
Download Citation
