A deep learning-based electric vehicle charging port recognition method and system

Abstract

This application relates to the field of image processing technology, and in particular to a method and system for recognizing electric vehicle charging ports based on deep learning. The method includes: acquiring an original image of the vehicle charging port and preprocessing it to obtain a pre-processed image; for any pixel, calculating its diffuse reflectance confidence score; weighting the gradient magnitude of the pixel based on the diffuse reflectance confidence score to obtain the optimal HOG feature; using a sliding window to perform a sliding search at each level of the pyramid in the pre-processed image, and calculating the charging port aggregation degree for any sliding window; adjusting the preset basic deformation weights in the DPM model scoring function based on the charging port aggregation degree to obtain the adaptive deformation penalty weights corresponding to the sliding window; calculating the comprehensive score of each sliding window using the DPM model scoring function based on the adaptive deformation penalty weights; and determining the charging port location based on the comprehensive score. This application can improve the accuracy of charging port recognition.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and system for recognizing electric vehicle charging ports based on deep learning. Background Technology

[0002] The charging port of an electric vehicle is the only physical interface connecting the power battery to the external power grid for energy transfer. With the development of autonomous driving technology and unattended charging stations, using vision systems to guide automated charging robots or robotic arms to automatically plug and unplug charging guns has become an industry trend. In this process, accurate identification and positioning of the charging port is a core prerequisite for ensuring the continuity and reliability of automated operations of robotic arms in complex outdoor environments.

[0003] To ensure operational efficiency on embedded devices, existing charging port identification technologies often employ the Deformable Part Model (DPM) algorithm. This algorithm decomposes the target into a root model and several part models for matching by calculating the Histogram of Gradient Orientations (HOG) features. It has the advantages of not relying on massive computing power and having a certain degree of robustness to local occlusion.

[0004] However, in real-world applications, vehicle paint surfaces are prone to specular reflection under strong light or supplemental lighting, creating bright spots and false edges. Simultaneously, background elements such as wheel rims and headlights often exhibit circular textures similar to those of the charging port. Because the original DPM algorithm primarily relies on gradient magnitude for feature extraction, and its component model deformation constraint parameters are fixed, the algorithm struggles to effectively distinguish between false edges caused by specular reflection and true physical edges when faced with strong reflective interference. Furthermore, it struggles to eliminate pseudo-targets with similar textures but loose structures when dealing with complex backgrounds. Ultimately, this leads to incorrect charging port identification in practical applications. Summary of the Invention

[0005] To improve the accuracy of charging port identification, this application provides a deep learning-based method and system for electric vehicle charging port identification.

[0006] Firstly, this application provides a deep learning-based method for identifying electric vehicle charging ports, employing the following technical solution:

[0007] A deep learning-based method for recognizing electric vehicle charging ports includes: acquiring an original image of the vehicle charging port and preprocessing it to obtain a preprocessed image; for any pixel, calculating its diffuse reflectance confidence, where the diffuse reflectance confidence is positively correlated with the pixel's saturation and negatively correlated with the pixel's brightness;

[0008] The gradient magnitude of pixels is weighted based on diffuse reflectance confidence to obtain the optimal HOG feature. A sliding window is used to perform a sliding search at each level of the pyramid of the initial image. For any sliding window, its charging port aggregation degree is calculated. The calculation steps of the charging port aggregation degree include: extracting multiple connected components in the image based on diffuse reflectance confidence, and taking the set of each connected component in the sliding window as the dark spot set; clustering the connected components in the dark spot set, and obtaining the charging port aggregation degree based on the area of ​​the connected components in each cluster.

[0009] The adaptive deformation penalty weight corresponding to the sliding window is obtained by adjusting the preset basic deformation weight in the DPM model scoring function based on the charging port aggregation degree. The adaptive deformation penalty weight is positively correlated with the charging port aggregation degree of the sliding window. The comprehensive score of each sliding window is calculated using the DPM model scoring function based on the adaptive deformation penalty weight, and the charging port position is determined based on the comprehensive score.

[0010] By calculating the diffuse reflectance confidence of each pixel and using this confidence to weight the gradient magnitude, interference from specular reflection is effectively suppressed at the feature extraction source. Simultaneously, the inherent porous connected domain topology of the charging port is used to calculate its aggregation degree, and the deformation penalty weight of the DPM model is adaptively adjusted accordingly. This achieves dynamic adjustment of the model's "rigidity" to the target shape, significantly improving the accuracy of charging port recognition under complex lighting and backgrounds, and effectively eliminating loosely structured false targets. Compared to the DPM algorithm in related technologies, this reduces background interference from textures similar to wheel hubs that are difficult to distinguish, thus improving the accuracy of charging port recognition.

[0011] Optionally, for any pixel, the step of calculating its diffuse reflectance confidence includes: taking the ratio of the pixel's saturation to its brightness as a first diffuse reflectance value; taking the difference between 1 and the normalized result of the pixel's brightness gradient magnitude as a second diffuse reflectance value; obtaining the diffuse reflectance confidence based on the first and second diffuse reflectance values, wherein the diffuse reflectance confidence is positively correlated with the first and second diffuse reflectance values.

[0012] For vehicle body paint, due to its reflective properties, pixels in this area typically exhibit high brightness and low saturation. Meanwhile, the charging port panel surface usually has a higher roughness, exhibiting diffuse reflection and balanced saturation and brightness. Therefore, the ratio of saturation to brightness, along with the normalized result of the brightness gradient amplitude, can indicate whether a pixel is likely in a high-brightness area, thus reducing interference from subsequent high-brightness regions.

[0013] Optionally, the product of the first diffuse reflectance value and the second diffuse reflectance value can be used as the diffuse reflectance confidence of the pixel.

[0014] By utilizing information from two dimensions—color distribution and edge sharpness—to construct the final confidence score, the accuracy of diffuse reflectance confidence score calculation is improved.

[0015] Optionally, the step of weighting the gradient magnitude of the pixel based on the diffuse reflection confidence to obtain the optimal HOG feature includes: for any pixel, taking the product of the diffuse reflection confidence and the gradient magnitude of the pixel as the optimal gradient, and obtaining the optimal HOG feature of the image based on the optimal gradient.

[0016] Traditional HOG feature calculations treat the gradient magnitude of all pixels equally, leading to the incorrect identification of false edges caused by reflections. This application directly applies diffuse reflection confidence as a weight to the gradient magnitude, reducing the contribution of gradient magnitude in highly reflective areas and generating optimal HOG features. This makes the feature descriptor more focused on the real object structure rather than lighting noise.

[0017] Optionally, in the step of clustering the connected domains in the dark spot set and obtaining the charging port aggregation degree based on the area of ​​the connected domains in each cluster, the number of clusters is 2, and two clusters are formed after clustering. The cluster with the largest mean area of ​​connected domains in the cluster is the main aperture cluster, and the other is the secondary aperture cluster. The charging port aggregation degree is positively correlated with the sum of the areas of connected domains in the main aperture cluster and negatively correlated with the standard deviation of the connected domains in both clusters.

[0018] Cluster analysis is used to classify dark spots into primary pore clusters and secondary pore clusters. The aggregation degree is constructed using the area mean and standard deviation. By taking advantage of the unique "DC hole + communication hole" structure of the charging port and the characteristic of multiple DC holes of the same size, it is possible to effectively distinguish simple circular interference objects from real charging ports and reduce the background false detection rate.

[0019] Optionally, the step of determining the charging port location based on the comprehensive score includes: sorting the comprehensive scores of each sliding window, using a non-maximum suppression algorithm to remove overlapping candidate windows, and retaining the candidate window with the highest score as the charging port location.

[0020] Optionally, the DPM model includes a root model and at least one component model, and the basic deformation weights are obtained offline based on the relative positions of the root model and the component model.

[0021] Optionally, the step of obtaining the adaptive deformation penalty weight corresponding to the sliding window based on the preset basic deformation weight in the DPM model scoring function according to the charging port aggregation degree includes: using the sum of 1 and the charging port aggregation degree as the adjustment coefficient, and using the product of the adjustment coefficient and the basic deformation weight as the adaptive deformation penalty weight.

[0022] By introducing the aggregation degree of the charging port as an adjustment coefficient into the deformation weight calculation, when a clear target that conforms to the topology of the charging port is detected, the deformation penalty weight is automatically increased, forcing the model to more strictly match the standard geometry, thereby enhancing the model's ability to reject pseudo-targets that are similar in shape but loosely structured.

[0023] Optionally, the product of the activation function value of the number of connected regions in the sliding window and the sum of the areas of the main aperture clusters is used as the first control value, the product of the standard deviations of the areas of the main aperture clusters and the secondary aperture clusters is used as the second control value, and the ratio between the first control value and the second control value is used as the charging port aggregation degree.

[0024] Secondly, this application provides a deep learning-based electric vehicle charging port identification system, which adopts the following technical solution:

[0025] A deep learning-based electric vehicle charging port identification system includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the system implements the aforementioned deep learning-based electric vehicle charging port identification method.

[0026] The aforementioned deep learning-based electric vehicle charging port identification method is used to generate a computer program, which is then stored in a memory for loading and execution by a processor. This allows for the creation of a system based on the memory and processor, making it convenient to use.

[0027] This application has the following technical advantages:

[0028] The diffuse reflection confidence score is constructed by utilizing the relationship between saturation and brightness. The gradient is then weighted based on the diffuse reflection confidence score to generate anti-reflective HOG features, which suppress false edges formed by the mirror reflection of the car paint. The aggregation index is constructed by utilizing the unique porous topology of the charging port, and the deformation penalty weight of the DPM model is adaptively adjusted to enhance the rigidity of the model when the features are clear, thereby eliminating false targets such as wheel hubs and improving the recognition accuracy of the charging port. Attached Figure Description

[0029] Figure 1 This is a flowchart of a method for identifying electric vehicle charging ports based on deep learning, according to an embodiment of this application.

[0030] Figure 2 This is a diffuse reflection confidence map in a deep learning-based electric vehicle charging port identification method according to an embodiment of this application.

[0031] Figure 3 This is a binarized dark spot region image after threshold segmentation in step S4 of a deep learning-based electric vehicle charging port identification method according to an embodiment of this application. Detailed Implementation

[0032] This application discloses a deep learning-based method for identifying electric vehicle charging ports. Starting from the reflection characteristics of physical optics, a diffuse reflection confidence map that reflects the material's reflectivity is constructed to suppress false edges. At the same time, the inherent porous topology of the charging port is utilized to construct a charging port aggregation index to eliminate background interference. Then, the aggregation index is used to adaptively adjust the model's penalty weight for component deformation, thereby achieving high-precision charging port identification and positioning under complex lighting and backgrounds.

[0033] This embodiment provides a deep learning-based method for recognizing electric vehicle charging ports. Starting from the reflection characteristics of physical optics, it constructs a light-material reflection difference map to suppress false edges. At the same time, it utilizes the inherent porous topology of the charging port to construct a charging port aggregation index to eliminate background interference. Then, it uses this aggregation index to adaptively adjust the model's penalty weight for component deformation, thereby achieving high-precision charging port recognition and positioning under complex lighting and backgrounds.

[0034] Reference Figure 1 A deep learning-based method for identifying electric vehicle charging ports includes steps S1-S6.

[0035] S1: Obtain the original image of the car charging port and perform preprocessing to obtain the initial processed image.

[0036] First, a high-precision image acquisition environment is constructed to obtain raw image data of the electric vehicle's charging port. When the electric vehicle enters the designated charging area and triggers a position sensor (such as an inductive loop or ultrasonic radar), the system controls the activation of a binocular RGB-D (Red Green Blue-Depth, infrared depth) camera mounted on the end effector of the automated charging robot. This camera integrates a CMOS (Complementary Metal Oxide Semiconductor) image sensor and an infrared ranging module, enabling it to acquire images of the charging port area, including the side charging port region, within a field of view of 0.5 to 2.0 meters. The acquired images are RGB images. To address the uncertainty of ambient lighting, the camera adjusts the exposure time using AEC (Automatic Exposure Control) and activates WDR (Wide Dynamic Range) mode, along with a side-mounted linear LED supplemental light, to enhance the detail of the dark holes inside the charging port. The image data is then wirelessly transmitted to the image processing unit of the industrial control computer.

[0037] Subsequently, the acquired raw images are preprocessed to eliminate environmental noise and enhance feature representation. Since sensor thermal noise and transmission noise in the raw images may interfere with gradient calculation, the color RGB images are first converted to grayscale images to reduce computational load. Next, since the charging port visually appears as a darker hole against a dark background, a morphological top-hat transform is used to process the grayscale image. This involves constructing a structuring element to perform an opening operation on the grayscale image, and then subtracting the opened image from the original grayscale image. This step corrects for uneven illumination and highlights the hole features in local dark areas. Finally, to eliminate high-frequency noise, a Gaussian filter is used to denoise the grayscale image after the top-hat transform. The Gaussian kernel size can be selected as 3×3 or 5×5, resulting in the preprocessed initial image.

[0038] S2: For any pixel, calculate its diffuse reflectance confidence level. The diffuse reflectance confidence level is positively correlated with the pixel's saturation and negatively correlated with the pixel's brightness.

[0039] In electric vehicle charging scenarios, the vehicle body paint is a smooth medium, resulting in specular reflection, characterized by high brightness and low saturation; while the charging port panel is a rough surface, producing diffuse reflection, with a relatively balanced ratio of saturation to brightness. To suppress false edge interference caused by vehicle body reflection, this step constructs a diffuse reflection confidence score that reflects the degree of diffuse reflection in the image.

[0040] First, the original RGB image before preprocessing in step one is converted to the HSV (Hue, Saturation, Value) color space, and the luminance channel V and saturation channel S are extracted. To eliminate the influence of sensor thermal noise on the values ​​of individual pixels, a 3×3 Gaussian smoothing filter is applied to both the V and S channels. Next, the Sobel operator is used to calculate the gradient magnitude of the luminance channel V, obtaining the gradient information of each pixel.

[0041] Based on the above information, for any pixel, the steps for calculating its diffuse reflectance confidence include: taking the ratio of the pixel's saturation to its brightness as the first diffuse reflectance value; taking the difference between 1 and the normalized result of the pixel's brightness gradient magnitude as the second diffuse reflectance value; obtaining the diffuse reflectance confidence based on the first and second diffuse reflectance values; and the diffuse reflectance confidence is positively correlated with the first and second diffuse reflectance values.

[0042] Specifically, for any pixel, the formula for calculating its diffuse reflectance confidence can be expressed as:

[0043] In the formula, coordinates The confidence level of diffuse reflection at a given point is normalized to [0,1]. coordinates The saturation of the pixels at that location; coordinates The brightness of the pixel at that location; To prevent the use of tiny constants with a denominator of zero, the value can be 0.001; coordinates The brightness gradient magnitude of the pixel at that location; This represents the maximum value of the brightness gradient magnitude of all pixels in the image.

[0044] Combination Figure 2 , This represents the first diffuse reflectance value, which utilizes the high brightness and low saturation characteristics of specular reflection areas. In highly reflective areas, Lower, and Larger, which leads to It approaches 0. However, this value is larger in the diffuse reflection area of ​​the charging port. Furthermore, reflective edges are usually accompanied by drastic gradient changes, i.e. Larger. This represents the second diffuse reflectance value at the edge of a strong gradient. Approaching 0. Multiplying the two results in a certain effect at the reflective areas and edges of the vehicle body. The value is extremely small, which allows us to distinguish false edges formed by reflections from the car paint.

[0045] S3: The gradient magnitude of the pixel is weighted based on the diffuse reflection confidence level to obtain the optimal HOG feature.

[0046] First, an image pyramid is constructed from the acquired images. When extracting HOG features for each layer, the diffuse confidence map obtained in step two is used. The gradient magnitudes are weighted. Specifically, the original gradient magnitude is multiplied by the gradient magnitude at the corresponding position to obtain the weighted optimal gradient, and then the HOG feature is calculated based on this to obtain the optimal HOG feature. This step purifies the features at the source and suppresses false gradients caused by reflections.

[0047] S4: Use a sliding window to perform a sliding search at each level of the pyramid of the initial processed image. For any sliding window, calculate its charging port aggregation degree.

[0048] Charging ports typically contain multiple dark-colored holes of uniform size and shape (such as DC terminal holes and communication holes), forming multiple dark spot areas. To utilize this topological feature to eliminate background interference, this step constructs the charging port aggregation degree of an arbitrary preset window.

[0049] Combination Figure 3First, the reflectance confidence map, composed of the diffuse reflectance confidence of each pixel, is used as input. Thresholding is then performed using Otsu's (Maximum Between-Class Variance Method) to obtain a binarized image. All connected components in the binarized image are extracted, and each connected component is treated as a dark spot.

[0050] Construct an image pyramid and traverse each layer of the pyramid using a sliding window of size 100×100 pixels. For the current layer and current position of the sliding window, count the set of dark spots that fall completely inside the window. For any dark spot in the set, the number of pixels in the dark spot is taken as the area of ​​the dark spot. Using the areas of all dark spots in the set as input data, the K-means (K-Means Clustering) algorithm is run, with the number of clusters set to 2, dividing the dark spots into two classes. The mean area of ​​the dark spots within each cluster is calculated, and the cluster with the larger mean is defined as the main aperture cluster. (Corresponding to larger DC apertures), clusters with smaller mean values ​​are defined as secondary aperture clusters. (Corresponding to a smaller communication port).

[0051] Based on the above classification results, calculate the charging port aggregation degree of the current sliding window. The formula for calculating the degree of aggregation of charging ports can be expressed as: ;

[0052] In the formula, The aggregation degree of the charging ports in the current sliding window; This represents the total number of dark spots contained within the window; The activation function is used to suppress cases where the number of dark spots is less than 2 (when...). When this value approaches 0 or is negative, the index value is suppressed. Indicates a main aperture cluster The sum of the areas of all dark spots within; The standard deviation of the dark spot area within the main aperture cluster; This represents the standard deviation of the dark spot area within the secondary pore cluster.

[0053] When the sliding window correctly frames the charging port, the window should contain multiple dark spots. Furthermore, after clustering, the main holes are similar in size, the secondary holes are similar in size, and the total area of ​​the main holes is relatively large. When these conditions are met, the aggregation degree of the charging port reaches its maximum. Conversely, if the frame represents background noise, the number of dark spots is usually small or their areas vary greatly, leading to... The value is very small.

[0054] S5: Based on the charging port aggregation degree, the preset basic deformation weight in the DPM model scoring function is adjusted to obtain the adaptive deformation penalty weight corresponding to the sliding window. The adaptive deformation penalty weight is positively correlated with the charging port aggregation degree of the sliding window.

[0055] The DPM algorithm allows components to displace relative to the root model to accommodate target deformation, but the charging port, as an industrial standard part, is rigid. To force the model to match standard geometry when features are clear, while maintaining a certain tolerance when features are ambiguous, this step constructs dynamic adaptive deformation penalty weights.

[0056] First, the DPM model was trained offline using the Latent SVM (Latent Support Vector Machine) algorithm. A positive sample set containing labeled charging port rectangles and a negative sample set without charging ports were constructed. The root filter was then trained using HOG (Histogram of Oriented Gradient) features. Based on the gradient energy distribution within the filter region, several regions with the highest energy are selected and initialized as component filters. The centers of these locations are used as the ideal anchor points for each component. The root filter is learned by iteratively optimizing and minimizing the loss function. Component filters and basic deformation weights .

[0057] For any sliding window, the formula for calculating its adaptive deformation penalty weight can be expressed as:

[0058] In the formula, This indicates the adaptive deformation penalty weights used when calculating the score function within the current sliding window; These are the basic deformation weights obtained through offline training. This refers to the aggregation degree of the charging port of the sliding window.

[0059] This represents the adjustment coefficient; the more obvious the detected structural features of the charging port, the higher the adjustment coefficient. The larger the value, The value increases accordingly. This means that the algorithm incurs a greater penalty for parts deviating from their ideal positions, forcing the model to become more "rigid," thus enabling it to forcefully eliminate false targets with similar textures but loose structures, such as wheel hubs. When features are not obvious, the model reverts to the basic deformation tolerance to avoid missed detections.

[0060] S6: Based on the adaptive deformation penalty weight, the DPM model scoring function is used to calculate the comprehensive score of each sliding window, and the charging port position is determined based on the comprehensive score.

[0061] For any sliding window, the formula for calculating its overall score can be expressed as:

[0062] In the formula, This represents the total score of the sliding window; The filter matching score for the root model; The number of components; For the first Filter matching score of each component model; The dynamic adaptive deformation penalty weight calculated for the current window; For the first The offset of each component relative to its ideal anchor point is calculated internally by the DPM model. This represents the side length of the current layer's mesh cell, primarily used for normalization. .

[0063] After completing the full image search, a series of candidate boxes with scores are obtained. The scores of all candidate boxes are sorted, and the NMS (Non-Maximum Suppression) algorithm is used to remove redundant boxes with high overlap. Finally, the window with the highest comprehensive score is retained as the recognition result of the charging port, and its position coordinates are output to guide the robotic arm to complete the automatic plugging and unplugging operation.

[0064] This application also discloses a deep learning-based electric vehicle charging port identification system, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a deep learning-based electric vehicle charging port identification method according to this application is implemented.

[0065] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0066] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A deep learning-based method for identifying electric vehicle charging ports, characterized in that, The original image of the car charging port is acquired and preprocessed to obtain the initial processed image; for any pixel, its diffuse reflectance confidence is calculated, which is positively correlated with the pixel's saturation and negatively correlated with the pixel's brightness. A sliding window is used to perform a sliding search at each level of the pyramid of the initial processed image. For any sliding window, its charging port aggregation degree is calculated. The steps for calculating the aggregation degree of the charging port include: extracting multiple connected components from the image based on diffuse reflectance confidence, taking the set of connected components in the sliding window as the dark spot set; clustering the connected components in the dark spot set, and obtaining the aggregation degree of the charging port based on the area of ​​the connected components in each cluster after clustering. The adaptive deformation penalty weight corresponding to the sliding window is obtained by adjusting the preset basic deformation weight in the DPM model scoring function based on the charging port aggregation degree. The adaptive deformation penalty weight is positively correlated with the charging port aggregation degree of the sliding window. The comprehensive score of each sliding window is calculated using the DPM model scoring function based on the adaptive deformation penalty weight, and the charging port position is determined based on the comprehensive score.

2. The method for identifying electric vehicle charging ports based on deep learning according to claim 1, characterized in that, For any pixel, the steps for calculating its diffuse reflectance confidence include: taking the ratio of the pixel's saturation to its brightness as the first diffuse reflectance value; taking the difference between 1 and the normalized result of the pixel's brightness gradient magnitude as the second diffuse reflectance value; obtaining the diffuse reflectance confidence based on the first and second diffuse reflectance values; and the diffuse reflectance confidence is positively correlated with the first and second diffuse reflectance values.

3. The method for identifying electric vehicle charging ports based on deep learning according to claim 2, characterized in that, The product of the first diffuse reflectance value and the second diffuse reflectance value is used as the diffuse reflectance confidence of the pixel.

4. The electric vehicle charging port identification method based on deep learning according to claim 1, characterized in that, The steps for obtaining the optimal HOG feature by weighting the gradient magnitude of pixels based on diffuse reflectance confidence include: for any pixel, the product of diffuse reflectance confidence and the gradient magnitude of the pixel is taken as the optimal gradient, and the optimal HOG feature of the image is obtained based on the optimal gradient.

5. The electric vehicle charging port identification method based on deep learning according to claim 1, characterized in that, In the step of clustering connected components in the dark spot set and obtaining the charging port aggregation degree based on the area of ​​connected components in each cluster, the number of clusters is 2, resulting in two clusters. The cluster with the largest mean area of ​​connected components is the main aperture cluster, and the other is the secondary aperture cluster. The charging port aggregation degree is positively correlated with the sum of the areas of connected components in the main aperture cluster and negatively correlated with the standard deviation of the connected components in both clusters.

6. The method for identifying electric vehicle charging ports based on deep learning according to claim 1, characterized in that, The steps for determining the charging port location based on the comprehensive score include: sorting the comprehensive scores of each sliding window, using a non-maximum suppression algorithm to remove overlapping candidate windows, and retaining the candidate window with the highest score as the charging port location.

7. The method for identifying electric vehicle charging ports based on deep learning according to claim 1, characterized in that, The DPM model includes a root model and at least one component model. The basic deformation weights are obtained offline based on the relative positions of the root model and the component model.

8. The electric vehicle charging port identification method based on deep learning according to claim 1, characterized in that, The steps for obtaining the adaptive deformation penalty weight corresponding to the sliding window based on the preset basic deformation weight in the DPM model scoring function based on the charging port aggregation degree include: using the sum of 1 and the charging port aggregation degree as the adjustment coefficient, and using the product of the adjustment coefficient and the basic deformation weight as the adaptive deformation penalty weight.

9. The method for identifying electric vehicle charging ports based on deep learning according to claim 5, characterized in that, The product of the activation function value of the number of connected regions in the sliding window and the sum of the areas of the main aperture clusters is used as the first control value, the product of the standard deviations of the areas of the main aperture clusters and the secondary aperture clusters is used as the second control value, and the ratio between the first control value and the second control value is used as the charging port aggregation degree.

10. A deep learning-based electric vehicle charging port identification system, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a deep learning-based electric vehicle charging port identification method according to any one of claims 1-9.

Images