Method for detecting vehicle fire in charging place

Abstract

The invention discloses a method for detecting vehicle fire in a charging place, which relates to the technical field of target identification, and is characterized by comprising the following steps of: inputting a shot target image of the charging place into a YOLOv5s-Fast model; a target image is processed through a YOLOv5s-Fast model, feature information of the target image is extracted, and an identification result is given. According to the method, on the basis of YOLOv5s, GC and C3 modules are fused in a Backbone network to form a module C3GC, so that the feature extraction capability of the model is enhanced, and the calculation overhead is reduced; a lightweight upsampling operator is adopted in the Neck network, so that the detection precision can be improved; a decoupling head is adopted to replace a coupling head in a Head network of YOLOv5s, so that the detection accuracy and efficiency are improved; compared with the quoted YOLOv5s, the method has the advantages that the average precision and the FPS are improved, and the method is more practical.

Description

Technical Field

[0001] The present invention belongs to the technical field of charging site detection, and in particular relates to a method for detecting vehicle fires in charging sites. Background Art

[0002] As an environmentally friendly and energy-saving means of transportation, electric vehicles are of great significance in reducing dependence on fossil fuels and reducing greenhouse gas emissions. With the increasing popularity of electric vehicles, charging stations are important facilities for their power supply. By the end of 2023, China will have built a total of 8.596 million charging infrastructure to meet the charging needs of electric vehicles. As the scale of charging facilities continues to expand, their safety issues are receiving more and more attention.

[0003] In recent years, vehicle charging fire accidents have emerged one after another, posing a serious threat to the life and health of people and the safety of property, and also affecting the normal operation of charging facilities. There are two main reasons for tram fires. First, the vehicle fire is caused by thermal runaway of the tram battery. Battery thermal runaway means that the internal temperature of the battery rises rapidly during use due to excessive heat, overcharging or short circuit, and eventually leads to a state of loss of control. This situation may cause accidents such as explosions and fires in tram batteries. Second, the vehicle fire is caused by the quality problems of the charging piles. Due to the uneven quality of domestic charging piles, the aging of the charging pile lines, the lack of overcharging protection, sensor failure and other problems can also cause charging vehicles to catch fire.

[0004] With the continuous advancement of computer vision technology in recent years, various target detection algorithms have also been widely used in fire monitoring and have made continuous progress. At present, although a variety of fire detection algorithms have been proposed, they still have defects in detection accuracy or speed. Moreover, some methods may not have a high detection and recognition rate in actual charging station scenarios; or some algorithms basically meet the requirements in terms of accuracy and speed, but the required computing power is too high and the practicality is low. These problems limit the application of existing technologies in actual charging station scenarios.

[0005] Therefore, developing an efficient target detection algorithm becomes particularly important for analyzing, predicting and responding to such accidents in a timely manner. Summary of the invention

[0006] The technical problem to be solved by the present invention is to solve the deficiencies in the prior art and design a new method for detecting vehicle fires in charging places. This detection method is based on YOLOv5s, and its Backbone network, Neck network and Head network are improved, which significantly improves the speed and accuracy of the model and solves the problem that the existing algorithm requires too much computing power and has low practicality.

[0007] The solution adopted by the present invention to solve the technical problem is:

[0008] A method for detecting a vehicle fire at a charging location, characterized in that it comprises the following steps:

[0009] Input the charging place target image obtained by the camera module into the YOLOv5s-Fast model;

[0010] The target image is recognized and processed by the YOLOv5s-Fast model to extract its feature information;

[0011] The recognition result is given according to the feature information.

[0012] As a preferred implementation of the present invention, the YOLOv5s-Fast model is based on YOLOv5s, and the GC and C3 modules are integrated in the Backbone network of YOLOv5s to form a new feature extraction module C3GC.

[0013] As a preferred implementation of the present invention, the YOLOv5s-Fast model is based on YOLOv5s and adopts a lightweight upsampling operator in the Neck network of YOLOv5s.

[0014] As a preferred implementation of the present invention, the YOLOv5s-Fast model is based on YOLOv5s, and a decoupling head is used instead of a coupling head in the Head network of YOLOv5s.

[0015] As a preferred embodiment of the present invention, the YOLOv5s-Fast model is based on YOLOv5s, and the GC and C3 modules are integrated in the Backbone network of YOLOv5s to form a new feature extraction module C3GC;

[0016] A lightweight upsampling operator is used in the Neck network of YOLOv5s;

[0017] In the Head network of YOLOv5s, the decoupling head is used instead of the coupling head.

[0018] The lightweight operator used is specifically the CARAFE operator.

[0019] The CARAFE operator is used to decompose the input feature map into a low-resolution feature map and a high-resolution feature map, upsample the low-resolution feature map to the same size as the high-resolution feature map by learning the relationship between pixels, and fuse the upsampled feature map with the high-resolution feature map to obtain the final upsampling result.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The present invention is based on YOLOv5s. Firstly, the global context attention mechanism (GC) is adopted in the Backbone network of YOLOv5s to be integrated with the C3 module to form a new feature extraction module C3GC, which enhances the model's ability to extract features and reduces the computational overhead; secondly, a lightweight upsampling operator is adopted in the Neck network of YOLOv5s, which can adaptively upsample according to the input image to improve the detection accuracy; finally, a decoupling head is adopted in the Head network of YOLOv5s to replace the coupling head to improve the accuracy and efficiency of target detection; or the three are used to improve the YOLOv5s model separately. Experimental and test results show that the method YOLOv5s-Fast proposed by the present invention has greatly improved average accuracy and FPS compared with the original YOLOv5s, and the method is more practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The network structure diagram of YOLOv5s cited in the present invention;

[0023] Figure 2 A schematic diagram of the network structure of the GC block cited in the present invention;

[0024] Figure 3 A comparison diagram of the network structures of C3GC and C3 in a method for detecting vehicle fires in charging locations proposed by the present invention;

[0025] Figure 4 This is a diagram of the CARAFE feature fusion structure cited in the present invention;

[0026] Figure 5 A structural comparison diagram of a coupling head and a decoupling head in a method for detecting a vehicle fire at a charging location proposed by the present invention;

[0027] Figure 6 This is a YOLOv5s-Fast algorithm structure diagram of the fourth embodiment of the method for detecting vehicle fires in charging locations proposed by the present invention. DETAILED DESCRIPTION

[0028] The specific implementation of the present invention is described below in conjunction with the accompanying drawings and embodiments:

[0029] It should be noted that the structures, colors, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so that people familiar with this technology can understand and read them, and are not used to limit the conditions under which the present invention can be implemented. Any structural modification, change in proportional relationship or adjustment of size should still fall within the scope of the technical content disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0030] At the same time, in the description of the present invention, it should be understood that the terms "one end", "the other end", "middle", "upper", "one side", "top", "inside", "front", "center", "both ends" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0031] In addition, the terms "first", "second", "third" and "fourth" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first", "second", "third" and "fourth" may explicitly or implicitly include at least one of such features.

[0032] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "screw-on" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0033] The present invention provides a method for detecting vehicle fire in a charging location, comprising:

[0034] Input the charging place target image obtained by the camera module into the YOLOv5s-Fast model;

[0035] The target image is recognized and processed by the YOLOv5s-Fast model to extract its feature information;

[0036] The recognition result is given according to the feature information.

[0037] In the first embodiment, the YOLOv5s-Fast model is based on YOLOv5s, and the GC and C3 modules are integrated in the Backbone network of YOLOv5s to form a new feature extraction module C3GC.

[0038] The YOLOv5 algorithm is a popular single-stage target detection algorithm at this stage. It has made significant improvements in speed and lightweight design. Its model is smaller and its practicality is stronger than later algorithms such as YOLOv7 and YOLOv8.

[0039] YOLOv5s is the smallest and fastest model in the YOLOv5 series. The network structure of YOLOv5s mainly consists of three parts: backbone network (Backbone), neck network (Neck), and head network (Head). Figure 1 shown.

[0040] The backbone network mainly includes the convolution module Conv, the C3 module and the fast spatial pyramid pooling module (Spatial Pyramid Pooling Fast, SPPF).

[0041] The convolution module is mainly composed of convolution layer, normalization layer and activation function, and its function is to extract target features. The C3 module is a unique module in the YOLOv5s algorithm, which uses residual learning and Concat to extract the main features. The SPPF module is improved on the SPP module. It uses multiple small pooling kernels to replace the single large pooling kernel of the SPP module, which improves the detection speed and can more completely retain the multi-scale information of the target during the degradation process, thereby optimizing the feature capture of multi-scale targets.

[0042] The neck network adopts the Feature Pyramid Networks (FPN) structure and is supplemented by the PANet structure to enhance feature fusion and positioning information transmission.

[0043] The head network is responsible for generating detection boxes and classifying, localizing, and scoring their confidence.

[0044] The GC block (Global Context Block) is an improvement on the non-local network. Experiments have shown that capturing long-distance dependencies is beneficial to various recognition tasks, while in traditional neural networks, long-distance dependencies can only be modeled by continuously stacking convolutional layers, which is inefficient and computationally intensive. For this reason, a GC block with better results is proposed.

[0045] Figure 2The network structure is a GC block. The input feature map is first calculated through a 1×1 convolution block and a Softmax module to obtain the attention weight feature information of the image, and then the global feature relationship information C×1×1 is calculated with the input H×W×C. Then, two 1×1 convolution blocks in the Transform structure are used to reduce the number of model parameters and save computational overhead. A LayerNorm module is inserted between the two 1×1 convolution blocks, which can improve the training stability of the model and has a certain regularization effect, reducing the risk of overfitting of the model. Finally, the initial information H×W×C is added to the global feature relationship information C×1×1 to obtain the output result of the enhanced global important information of the image.

[0046] Figure 3 The new module C3GC proposed in this paper is compared with the existing C3 module in detail. In the Backbone part, as the number of network layers increases, the C3 module faces challenges in extracting shallow features and is prone to the problem of gradient vanishing when the network depth increases. In view of this, this paper cleverly embeds the GC block before the two Conv modules. The GC block effectively promotes the model's ability to extract global context features before the Conv module through global modeling relationships, thereby improving the accuracy of the model and reducing computational overhead.

[0047] In the second embodiment, the YOLOv5s-Fast model is based on YOLOv5s and adopts a lightweight upsampling operator in the Neck network of YOLOv5s.

[0048] In the YOLOv5s algorithm, the Neck part mainly performs feature amplification and fusion on the high-level feature information extracted by Backbone. The purpose of upsampling is to restore the image to its original size so that the model can retain as much information as possible when the network depth increases. In YOLOv5s, the upsampling method used is the nearest neighbor interpolation method, which uses the grayscale values ​​of adjacent pixels as the grayscale values ​​of newly generated pixels. The disadvantage of this method is that the image is prone to information loss and jagged edges, resulting in reduced accuracy. Therefore, this paper adds a new lightweight upsampling operator (Content-Aware ReAssembly Feature Extraction, CARAFE) to replace the upsampling method in YOLOv5s to improve the model's feature fusion capabilities.

[0049] CARAFE is an upsampling operator. Its core idea is to sample by learning the relationship between pixels and to perform weighted fusion on surrounding pixels, thereby obtaining rich contextual features. Compared with the neighbor interpolation method used in YOLOv5s, the CARAFE operator has a larger receptive field than the neighbor interpolation method, which can obtain more contextual feature information and improve detection accuracy. Secondly, CARAFE only takes up very little computational overhead and can be integrated into several mainstream network architectures, making CARAFE a lightweight and efficient operator suitable for scenarios with limited computing resources.

[0050] Figure 4 This is the feature fusion structure diagram of CARAFE. The implementation method of the CARAFE operator is to decompose the input feature map into two parts, namely the low-resolution feature map and the high-resolution feature map. Then, by learning the relationship between pixels, the low-resolution feature map is upsampled to the same size as the high-resolution feature map. Finally, the upsampled feature map is fused with the high-resolution feature map to obtain the final upsampling result. This method helps to improve the quality and fidelity of sampling and is suitable for tasks that require higher quality image processing.

[0051] In the third embodiment, the YOLOv5s-Fast model is based on YOLOv5s, and a decoupling head is used instead of a coupling head in the Head network of YOLOv5s.

[0052] The head network of YOLOv5s uses a coupling head design, which is essentially a technology that fuses feature maps of different levels and jointly predicts the target. However, this network design will increase the amount of calculation and complexity of the model. Secondly, it requires a large number of data sets to train to achieve the best results, which will cause the risk of overfitting for small data sets. Therefore, this paper uses a decoupling head to replace the coupling head in YOLOv5s.

[0053] The decoupling head is a head network design in the field of target detection. Its design idea is to divide the target detection task into two parts: classification and regression. This design idea can improve the accuracy and efficiency of the target detection task.

[0054] Figure 5 That is, the structural comparison between the coupling head and the decoupling head. As shown in the figure, the coupling head only reduces the channel of the input feature map to 33 through a 1×1 convolution and then directly performs classification and regression. This will create the risk of model overfitting. The decoupling head first reduces the channel to 128 through a 1×1 convolution, and then divides it into two branches to perform classification and regression through two 3×3 convolutions respectively. In comparison, the decoupling head can retain more information and improve the accuracy.

[0055] In the fourth embodiment, the YOLOv5s-Fast model is based on YOLOv5s, and the GC and C3 modules are integrated in the Backbone network of YOLOv5s to form a new feature extraction module C3GC;

[0056] A lightweight upsampling operator is used in the Neck network of YOLOv5s;

[0057] In the Head network of YOLOv5s, the decoupling head is used instead of the coupling head.

[0058] The lightweight upsampling operator is specifically the CARAFE operator.

[0059] Figure 6 This is the YOLOv5s-Fast structure diagram in the fourth embodiment. In the Backbone part, the four C3 modules are replaced by the C3GC module proposed by the present invention after being integrated with the global context attention mechanism; in the Neck part, the present invention changes the upsampling operator to the lightweight operator CARAFE; in the Head part, the original coupling head is replaced by the decoupled head.

[0060] The CARAFE operator is used to decompose the input feature map into a low-resolution feature map and a high-resolution feature map. By learning the relationship between pixels, the low-resolution feature map is upsampled to the same size as the high-resolution feature map, and the upsampled feature map is fused with the high-resolution feature map to obtain the final upsampling result.

[0061] In order to verify the performance of the algorithm proposed in the present invention, the vehicle fire dataset in the public Traffic-Net dataset is used for calculation, with a total of 1006 vehicle fire images. The dataset is labeled using the LabelImg tool, and the labels are in YOLO format. There are three labels: car, fire, and smoke, where car represents the car in the image, fire represents the fire point, and smoke represents smoke. The labeled dataset is divided into a training set and a test set at a ratio of 8:2. The training set, test set, and the number of various labels are shown in Table 1.

[0062] Table 1 Data sample

[0063]

[0064] The experimental environment of the algorithm proposed in this invention is based on Windows 10 operating system, using NVIDIA GeForce GTX 1650 as the hardware graphics card, GPU version of Pytorch as the deep learning framework, Python version 3.8.2, CUDA computing library version 12.1, and the number of training rounds epoch is set to 100. Among them, the momentum is 0.937, the weight decay is 0.0005, the batch size is 8, the learning rate is 0.0001, and the input image size is 640×640.

[0065] Precision (P), recall (R) and mean average precision (mAP) are three key indicators for evaluating model performance.

[0066]

[0067]

[0068]

[0069]

[0070] In the above formula, precision P represents the proportion of true examples in the predicted samples, and recall R, also known as recall rate, represents the proportion of positive examples in the sample to the predicted samples. P Indicates the number of samples correctly predicted as positive, F P Expressed as the number of samples that are incorrectly predicted to be positive; F N It is expressed as the number of negative samples that the model incorrectly predicts. The average precision (AP) is obtained by plotting the PR curve and calculating the area under it. mAP is the average of the AP values ​​of all categories and is used to measure the overall performance.

[0071] In addition, the indicator for measuring the detection speed of the model is the number of frames per second (FPS). FPS represents the number of images processed per second. The higher the FPS, the faster the detection speed of the model. The indicator for measuring the complexity of the model is the number of floating point operations per second (GFLOPs). It represents the number of floating point operations performed per second and is often used to measure the computational time complexity (i.e., the amount of computation) of the model.

[0072] In order to verify the effectiveness of the model after adding the C3GC module, replacing the CARAFE upsampling operator, and replacing the detection head, this paper conducted an ablation experiment and compared the results. The specific experimental data are shown in Table 2. Model 0 is the YOLOv5s algorithm; Model 1 is the YOLOv5s+C3GC module (i.e., the first embodiment); Model 2 is the YOLOv5s+CARAFE upsampling operator (i.e., the second embodiment); Model 3 is the YOLOv5s+decoupling head (i.e., the third embodiment); Model 4 is the final improved algorithm after integrating the above three improvements (i.e., the fourth embodiment). It can be seen from Table 2 that the recall rate R of the YOLOv5s algorithm, i.e., Model 0, is significantly low when detecting vehicle fires, which proves that the unimproved YOLOv5s algorithm has many missed detection phenomena. Compared with model 0, the mAP@0.5 / % of models 1, 2, and 3 have been improved to varying degrees, especially the recall rate R has been improved significantly; model 4, which is the final improved algorithm after integrating the three improvements, has a mAP@0.5 / % higher than the original YOLOv5s algorithm by 4.9%, and the FPS has increased from 46.225 to 59.462. In addition, indicators such as precision and recall have also been improved to varying degrees, proving the effectiveness and practicality of the improved YOLOv5s-Fast algorithm.

[0073] Table 2 Comparison of ablation experiment results

[0074] Model +C3GC +CARAFE +Decoupling head P / ％ R / ％ mAP@0.5 / % FPS 0 72.3 54.3 61.0 46.225 1 √ 67.6 61.7 63.8 55.623 2 √ 66.9 62.2 61.1 49.289 3 √ 67.9 58.7 61.9 45.035 4 √ √ √ 68.1 62.1 65.9 59.462

[0075] Since the research goal of this paper is to detect vehicle fires in charging stations, the smoke and flames around the vehicle are regarded as targets to be detected. This detection mechanism ensures that the model can detect smoke or flames around the vehicle in a timely manner, thereby improving the safety of vehicle charging.

[0076] The YOLOv5s algorithm does not perform well when detecting feature maps of different sizes and resolutions that are cropped from the video. A large number of missed detections occur, and the confidence level of the detected labels is also low. Moreover, the detection results of the YOLOv5s algorithm for the four images of different resolutions and sizes are also very different.

[0077] However, the algorithm proposed in the present invention avoids these problems very well. In terms of recognition accuracy, firstly, thanks to the application of the C3GC module in the Backbone part, the algorithm of this paper is more comprehensive in extracting contextual features. Secondly, thanks to the use of the lightweight upsampling operator CARAFE, it performs adaptive upsampling by learning the contextual relationship between images, while the neighboring interpolation rule used in the upsampling of YOLOv5s is to calculate the pixel value to be restored by the pixel values ​​of several adjacent points. This method will cause a large amount of information loss in the upsampling process of YOLOv5s. Therefore, the algorithm of the present invention is significantly better than the YOLOv5s algorithm in practical applications, especially when processing high-pixel and high-definition images. Thirdly, the present invention replaces the head network part with a decoupling head, which processes the classification and regression tasks separately. Compared with the coupling head in YOLOv5s, the decoupling head retains more information and improves the accuracy of the model.

[0078] In terms of reasoning speed, the average time taken by YOLOv5s to reason about four images is 28ms, while the average time taken by the algorithm proposed in the fourth embodiment of the present invention to reason about four images is only 16ms. Different from the YOLOv5s algorithm that transmits extracted information by stacking convolutional layers, the faster reasoning speed of the algorithm proposed in the present invention is also due to the use of the C3GC module and the lightweight upsampling operator, which promotes the model to fully extract context features and reduces the amount of calculation.

[0079] In terms of model complexity, the GFLOPs of YOLOv5s is 16.1, while the GFLOPs of the algorithm YOLOv5s-Fast proposed in the fourth embodiment of the present invention is 16.2. GFLOPs is an important indicator for measuring model complexity and computational complexity. It can be seen that there is only a slight difference in model complexity between the two.

[0080] In summary, it is proved that the algorithm proposed in the present invention is superior to the YOLOv5s algorithm in terms of practicality and generalization, and is superior to the YOLOv5s algorithm in terms of accuracy and inference speed when processing images, which proves the effectiveness of the improved algorithm of the present invention.

[0081] In order to verify the superiority of the YOLOv5s-Fast model of the present invention, it is compared with a variety of mainstream detection models for the vehicle fire data set, as shown in Table 3. First, the Faster R-CNN algorithm has the highest mAP@0.5, but the lowest FPS, because it is a two-stage target detection algorithm, so the detection speed is slow; secondly, the YOLOv7, YOLOv8, YOLOv9 and YOLOv5s algorithms are all single-stage algorithms with fast detection speeds, and the mAP@0.5 and FPS of the four YOLO algorithms are not much different, but the model size of YOLOv5s is smaller than the above three YOLO algorithms, that is, YOLOv5s only needs to occupy less computing resources when actually deployed. Therefore, the comprehensive comparison shows that the YOLOv5s algorithm is more suitable for application in the actual industrial field, so this paper chooses to improve on the basis of the YOLOv5s algorithm. The experimental data proves that the algorithm YOLOv5s-Fast proposed in this invention has higher detection speed and accuracy than YOLOv5s on the premise that the model complexity is not much different, and can well meet the requirements of vehicle fire detection in charging stations.

[0082] Table 3 Model performance comparison

[0083]

[0084]

[0085] The present invention proposes the YOLOv5s-Fast algorithm for the specific scenario of vehicle fire detection in charging stations. The YOLOv5s-Fast algorithm is based on YOLOv5s, introduces the global context attention mechanism and the original C3 module fusion, and introduces the lightweight upsampling operator CARAFE and the decoupling head, or improves the three separately.

[0086] Experimental results show that, compared with the YOLOv5s algorithm, the fourth embodiment of YOLOv5s-Fast proposed in the present invention improves the average accuracy by 4.9% and the FPS from the original 46 to 59 while keeping the model complexity almost the same, which significantly improves the speed and accuracy of the model.

[0087] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

[0088] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.

Claims

1. A method for detecting vehicle fires at a charging location, characterized in that: Include: Input the charging place target image obtained by the camera module into the YOLOv5s-Fast model; The target image is recognized and processed by the YOLOv5s-Fast model to extract its feature information; The recognition result is given according to the feature information.

2. A method for detecting vehicle fire in a charging location as claimed in claim 1, characterized in that: The YOLOv5s-Fast model is based on YOLOv5s, and integrates the GC and C3 modules in the Backbone network of YOLOv5s to form a new feature extraction module C3GC.

3. A method for detecting vehicle fire in a charging location as claimed in claim 1, characterized in that: The YOLOv5s-Fast model is based on YOLOv5s and uses a lightweight upsampling operator in the Neck network of YOLOv5s.

4. A method for detecting vehicle fire in a charging location as claimed in claim 1, characterized in that: The YOLOv5s-Fast model is based on YOLOv5s, and a decoupling head is used instead of a coupling head in the Head network of YOLOv5s.

5. A method for detecting vehicle fire in a charging location as claimed in claim 1, characterized in that: The YOLOv5s-Fast model is based on YOLOv5s, and integrates the GC and C3 modules in the Backbone network of YOLOv5s to form a new feature extraction module C3GC; A lightweight upsampling operator is used in the Neck network of YOLOv5s; In the Head network of YOLOv5s, the decoupling head is used instead of the coupling head.

6. A method for detecting vehicle fire in a charging location according to any one of claims 3 or 5, characterized in that: The lightweight operator used is specifically the CARAFE operator.

7. A method for detecting vehicle fire in a charging location as claimed in claim 6, characterized in that: The CARAFE operator is used to decompose the feature map after upsampling on the Neck network of YOLOv5s into a low-resolution feature map and a high-resolution feature map, upsample the low-resolution feature map to the same size as the high-resolution feature map by learning the relationship between pixels, and fuse the upsampled feature map with the high-resolution feature map to obtain the final upsampling result.

Images