Roadside intelligent parking target detection method based on improved YOlOv8 model

Abstract

The invention provides a roadside intelligent parking target detection method based on an improved YOlOv8 model, and the method comprises the steps: collecting a roadside parking image, carrying out the preprocessing and marking of the roadside parking image, generating a roadside parking data set, carrying out the improvement of a conventional YOLOv8 model, building a roadside intelligent parking target detection model based on the improved YOLOv8 model, and carrying out the detection of the roadside intelligent parking target. And training the roadside intelligent parking target detection model based on the roadside parking data set to obtain a final roadside intelligent parking target detection model, and inputting the to-be-detected image into the final roadside intelligent parking target detection model to obtain a detection result. According to the roadside parking detection method, the YOLOv8 model is subjected to structure optimization and lightweight processing, so that the calculated amount and power consumption are remarkably reduced on the premise of ensuring the detection precision, and low-cost, efficient and real-time roadside parking detection is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation systems, and in particular to a roadside intelligent parking target detection method based on an improved YOlOv8 model. Background Art

[0002] With the acceleration of urbanization, the demand for parking on urban roads is increasing, and the demand for intelligent parking management is becoming stronger. The existing target detection model has the problems of low recognition accuracy and high complexity in detecting roadside vehicles.

[0003] The YOLO (You Only Look Once) series of models are object detection algorithms that have developed rapidly in recent years. The YOLOv8 model, as the latest generation of YOLO models, has excellent performance in detection accuracy and speed. However, direct application to embedded devices has problems such as high computational complexity and high power consumption.

[0004] Therefore, it is necessary to design a roadside intelligent parking target detection method based on the improved YOlOv8 model. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the object of the present invention is to provide a roadside intelligent parking target detection method based on an improved YOlOv8 model.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A roadside smart parking target detection method based on an improved YOlOv8 model, comprising:

[0008] Collect roadside parking images, preprocess and annotate them, and generate roadside parking datasets;

[0009] Improve the traditional YOLOv8 model and build a roadside smart parking target detection model based on the improved YOLOv8 model;

[0010] The roadside smart parking target detection model is trained based on the roadside parking dataset to obtain the final roadside smart parking target detection model;

[0011] The image to be detected is input into the final roadside intelligent parking target detection model to obtain the detection result.

[0012] Optionally, collect roadside parking images, and pre-process and annotate them, specifically:

[0013] Collect a large number of roadside parking images at urban roadside parking spaces based on image acquisition equipment;

[0014] Perform data enhancement processing on the collected roadside parking images;

[0015] Annotate the enhanced roadside parking images, annotate the bounding box of each vehicle, the bounding box of the parking space, the license plate number of each vehicle, and the parking space number of the parking space;

[0016] Generate a roadside parking dataset based on the processed images.

[0017] Optionally, data enhancement processing is performed on the collected roadside parking images, specifically:

[0018] Data enhancement processing is performed on the collected roadside parking images based on the DM-GAN network model;

[0019] Perform regional image enhancement on the roadside parking image after data enhancement.

[0020] Optionally, data enhancement processing is performed on the collected roadside parking images based on the DM-GAN network model, specifically:

[0021] Construct DM-GAN network model;

[0022] The collected roadside parking images are enhanced based on the DM-GAN network model.

[0023] Optionally, a DM-GAN network model is constructed, specifically:

[0024] Construct the DM-GAN network model, including the generator network module, the discriminator module and the loss function module;

[0025] The generator network module is based on the U-Net network as the backbone, and the DC dual-channel attention mechanism is integrated into the jump connection part of the U-Net network. The MSDR multi-scale dense residual block is introduced at the bottom of the generator network module. The MSDR multi-scale dense residual block is designed by combining multi-scale feature extraction and dense residual.

[0026] The discriminator module consists of the Patch-GAN network architecture;

[0027] The loss function module includes perceptual loss, adversarial loss and SF loss.

[0028] Optionally, the roadside parking image after data enhancement processing is subjected to regional image enhancement processing, specifically:

[0029] Obtain a roadside parking image after data enhancement processing, which is an RGB format image;

[0030] The roadside parking image after data enhancement is converted from RGB format to LAB format, and an L component image is obtained;

[0031] Calculate the pixel mean of all pixels on the L component image to obtain the L component mean;

[0032] Perform Gaussian smoothing on the L component image to obtain a Gaussian L component image;

[0033] The pixel difference between each pixel point on the Gaussian L component image and the L component mean is calculated in sequence, and the salient area image is obtained according to the size of the pixel difference corresponding to each pixel point;

[0034] Normalizing the salient region image to obtain a normalized image;

[0035] Binarize the normalized image to obtain a binary image;

[0036] Extract edge information of the binary image to obtain an edge image;

[0037] The binary image and edge image are used as input, and the connected region detection method is used to extract the connected region of the normalized image, and the connected region is the region to be enhanced;

[0038] An image enhancement algorithm is used to enhance all pixels in the enhanced area to obtain an image after regional image enhancement.

[0039] Optionally, the traditional YOLOv8 model is improved as follows:

[0040] Faster Block is designed based on PConv in FasterNet. In the feature extraction C2f module of the traditional YOLOv8 model, Faster Block replaces Bottleneck as the main gradient flow branch by class inheritance.

[0041] Replace the feature fusion network of the traditional YOLOv8 model with the BiFPN network structure;

[0042] The dynamic non-monotonic focusing mechanism WIoU loss function is selected to replace the loss function of the traditional YOLOv8 model.

[0043] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0044] The present invention provides a roadside smart parking target detection method based on an improved YOLOv8 model, the method comprising: collecting roadside parking images, preprocessing and labeling them, generating a roadside parking data set, improving the traditional YOLOv8 model, and constructing a roadside smart parking target detection model based on the improved YOLOv8 model, training the roadside smart parking target detection model based on the roadside parking data set, obtaining a final roadside smart parking target detection model, inputting the image to be detected into the final roadside smart parking target detection model, and obtaining a detection result. The present invention optimizes the structure and lightweights the YOLOv8 model, so that the amount of calculation and power consumption are significantly reduced under the premise of ensuring detection accuracy, and low-cost, efficient, and real-time roadside parking detection is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0046] Figure 1 Schematic diagram of the DM-GAN network model structure;

[0047] Figure 2 Schematic diagram of DC dual channel attention mechanism;

[0048] Figure 3 This is a schematic diagram of the MSDR network structure;

[0049] Figure 4 This is a schematic diagram of the improved YOLOv8 model structure;

[0050] Figure 5 This is a schematic diagram of the Bottleneck structure;

[0051] Figure 6 It is a schematic diagram of the C2f-PConv module structure;

[0052] Figure 7a Schematic diagram of the original BiFPN structure;

[0053] Figure 7b To improve the BiFPN structure diagram;

[0054] Figure 8 Schematic diagram of the intersection area of ​​the real box and the predicted box;

[0055] Fig. 9 A flow chart of a method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The purpose of the present invention is to provide a roadside intelligent parking target detection method based on an improved YOLOv8 model. By optimizing the structure and lightweighting the YOLOv8 model, the calculation amount and power consumption are significantly reduced while ensuring the detection accuracy, thereby achieving low-cost, efficient and real-time roadside parking detection.

[0058] Fig. 9 A flow chart of a method provided by an embodiment of the present invention, such as Fig. 9 As shown, the present invention provides a roadside smart parking target detection method based on an improved YO10v8 model, comprising:

[0059] Step 100: Collect roadside parking images, preprocess and annotate them, and generate a roadside parking dataset;

[0060] Step 200: improving the traditional YOLOv8 model, and building a roadside smart parking target detection model based on the improved YOLOv8 model;

[0061] Step 300: training a roadside smart parking target detection model based on a roadside parking dataset to obtain a final roadside smart parking target detection model;

[0062] Step 400: Input the image to be detected into the final roadside smart parking target detection model to obtain the detection result.

[0063] Step 100 specifically includes the following steps:

[0064] Step 101: Collecting a large number of roadside parking images at roadside parking spaces in the city based on an image acquisition device;

[0065] Step 102: performing data enhancement processing on the collected roadside parking image;

[0066] Step 103: annotating the roadside parking image after data enhancement, annotating the bounding box of each vehicle, the bounding box of the parking space, the license plate number of each vehicle, and the parking space number of the parking space;

[0067] Step 104: Generate a roadside parking dataset based on the processed images.

[0068] The specific process in step 102 includes the following steps:

[0069] Step 1021: performing data enhancement processing on the collected roadside parking images based on the DM-GAN network model;

[0070] Step 1022: performing regional image enhancement processing on the roadside parking image after data enhancement processing.

[0071] The specific process in step 1021 includes the following steps, and each step is expanded:

[0072] S1. Construct the DM-GAN network model, specifically:

[0073] The DM-GAN network model structure is as follows Figure 1 As shown, it includes a generator network module, a discriminator module and a loss function module;

[0074] Among them, the generator network module is introduced:

[0075] The generator network module is based on the U-Net network as the backbone, and the DC dual channel attention mechanism is integrated into the jump connection part of the U-Net network. Figure 2 As shown in the figure, it is divided into two branches, the upper part introduces average pooling and maximum pooling, which are used to aggregate the spatial information of feature maps between channels to better capture the texture and background information of the image. The lower part is the SE (squeeze and excitation) module, which is used to adaptively learn channel weights to enhance useful feature channels. Through the complementary learning of the two parts, the network can focus more on the detailed texture, color and other information of the image;

[0076] The MSDR multi-scale dense residual block is introduced at the bottom layer of the generator network module. The MSDR multi-scale dense residual block is designed by combining multi-scale feature extraction and dense residual. The MSDR network structure is as follows: Figure 3 As shown in the figure, the number of channels is first divided into 4 subsets, and convolution kernel sizes of 1×1, 3×3, 5×5, and 7×7 are used to obtain multi-scale receptive fields. Then, the feature mapping method is used to enable each convolution layer to accept and fuse multi-scale feature information from the previous layer to obtain different receptive fields, which is conducive to feature reuse and reduces image feature loss. At the same time, the use of residual modules helps alleviate the problem of gradient dispersion and makes training faster;

[0077] The discriminator module adopts the Patch-GAN network architecture, and its output is an N×N matrix. The mean of the matrix is ​​used to judge the probability that the image is a real image. It not only considers the consistency of the entire image, but also pays attention to the local details of the image. Figure 1It can be seen that the discriminator D is composed of 4 fully convolutional layers with 64, 128, 256 and 512 channels respectively, which are used to extract the feature information of the image layer by layer. Compared with the global discriminator, the local discriminator requires fewer parameters and is more efficient;

[0078] The loss function module includes perceptual loss, adversarial loss and SF loss, which are introduced as follows:

[0079] (1) Perceptual loss

[0080] Perceptual loss is used to limit the feature extraction distance between the image generated by the generator and its corresponding target image. Using the VGG perceptual loss function can reduce the high-frequency feature loss caused by the network during the encoding and decoding process. The perceptual loss formula is as follows:

[0081]

[0082] Where: I C represents the target image, I D represents the image generated by the generator, C j H j W j refers to the size of the feature map of the jth layer, and φ is the high-level semantic information extracted by the VGG19 network.

[0083] (2) Fighting Losses

[0084] Adversarial loss can not only optimize the network training process, but also guide the generator to generate more realistic images, improve the discriminator's ability to distinguish authenticity, and ensure that the generated image and the target image have the same style. The adversarial loss function is as follows:

[0085] L adv (D,G)=E x-p(x) [logD(x)]+E z-p(z) [log(1-D(G(z)))]

[0086] Where, L adv (D, G) is the result of continuous updating and optimization of the generator and discriminator. G is the generator network, D is the discriminator network, E represents mathematical expectation, x refers to the input image, and z represents noise.

[0087] (3) SF loss

[0088] In order to reduce the difference between pixels and retain the integrity of image edge information, and avoid exposure and blur in the enhanced image, SF loss starts from two aspects: image deblurring and detail enhancement. On the one hand, it reduces noise and blur in the image by calculating the average of the absolute difference between the network output image and the real image, but it will cause the edge of the image to be too smooth and lose some information; on the other hand, it smoothes the image by measuring the difference between adjacent pixels in the image, retaining the structural features and edge information of the image. The two are allocated and combined according to certain weights. The SF loss function formula is as follows:

[0089]

[0090] In the formula, x represents the input image, y represents the output image, C, H and W represent the generated image Kω(I s ), and They represent the gradients in the x-direction and y-direction respectively. Through continuous learning of the network, λ1=λ2=0.5 is finally determined.

[0091] The total network loss is a linear combination of the three, and the formula is as follows:

[0092] L total =I adv +L con +L SF

[0093] S2. Perform data enhancement processing on the collected roadside parking images based on the DM-GAN network model.

[0094] The specific process in step 1022 includes:

[0095] Obtain a roadside parking image after data enhancement processing, which is an RGB format image;

[0096] The roadside parking image after data enhancement is converted from RGB format to LAB format, and an L component image is obtained;

[0097] Calculate the pixel mean of all pixels on the L component image to obtain the L component mean;

[0098] Perform Gaussian smoothing on the L component image to obtain a Gaussian L component image;

[0099] The pixel difference between each pixel point on the Gaussian L component image and the L component mean is calculated in sequence, and the salient area image is obtained according to the size of the pixel difference corresponding to each pixel point;

[0100] Normalizing the salient region image to obtain a normalized image;

[0101] Binarize the normalized image to obtain a binary image;

[0102] Extract edge information of the binary image to obtain an edge image;

[0103] The binary image and edge image are used as input, and the connected region detection method is used to extract the connected region of the normalized image, and the connected region is the region to be enhanced;

[0104] An image enhancement algorithm is used to enhance all pixels in the enhanced area to obtain an image after regional image enhancement.

[0105] In step 200, the improvement process of the traditional YOLOv8 model specifically includes:

[0106] FasterBlock is designed based on PConv in FasterNet. In the feature extraction C2f module of the traditional YOLOv8 model, FasterBlock replaces Bottleneck as the main gradient flow branch by class inheritance.

[0107] Replace the feature fusion network of the traditional YOLOv8 model with the BiFPN network structure;

[0108] Select the dynamic non-monotonic focusing mechanism WIoU loss function to replace the loss function of the traditional YOLOv8 model;

[0109] The above content is introduced in detail. The improved YOLOv8 model is as follows Figure 4 As shown;

[0110] Regarding the first improvement point:

[0111] The C2f module used by YOLOv8 contains more Bottleneck structures, which leads to excessive redundancy of channel information while extracting more features. At present, lightweight networks use deep convolution or group convolution to extract spatial features. While reducing the amount of calculation, the operator causes an increase in memory access and low efficiency of fragmented calculation. Among them, PConv in FasterNet uses the redundant information in the feature map and only performs regular convolution on some channels without affecting other channels. This can better utilize the computing power of the device and reduce computational redundancy and memory access. PConv only needs to perform convolution operations on some channels of the input to extract spatial features, while keeping other channels unchanged. Assume that h and w represent the length and width of the input feature map, c represents the number of input channels, C represents the channels participating in the convolution, k is the convolution kernel size, r represents the rate of participation in the convolution, and its computational amount F PCony Expression, memory access amount M AC The calculation formula is as follows:

[0112]

[0113]

[0114] The typical convolution rate is 1 / 4, and the FLOPs of PConv is only 1 / 16 of that of conventional convolution. In addition, since the convolution process occupies less memory, the memory access volume is about 1 / 4 of that of conventional convolution.

[0115] Since only C n The channel participates in spatial feature extraction, so there is no loss of subsequent feature channel information, which can greatly reduce the amount of calculation and memory access.

[0116] The series-connected Bottleneck modules can extract and fuse features of different scales and improve the representation capability of feature maps. However, since YOLOv8 contains a large number of Bottleneck structures, it brings greater computational overhead. Figure 5 .

[0117] The present invention uses PCony in FasterNet to design FasterBlock. In the feature extraction C2f module, Faster Block replaces Bottleneck as the main gradient flow branch by class inheritance, reducing the model floating point number and calculation amount. By constructing C2f-PConv and introducing PConv, the floating point number and calculation amount in the convolution process are further reduced. The C2f-PConv module structure is as follows Figure 6 shown.

[0118] Improvements for the second point:

[0119] Multi-scale feature fusion aims to aggregate features of different resolutions. PANet has two paths, bottom-up and top-down. The effect is better than FPN, but the computational complexity is larger. BiFPN is based on PANet and removes nodes with only one input. Since such nodes do not perform feature fusion, their contribution to the feature fusion network is small. At the same time, in order to fuse more features, an additional path is added to connect the input nodes and output nodes of the same-level network. The original BiFPN structure is as follows: Figure 7a As shown, the present invention improves the BiFPN structure by adding a P2 large-size feature layer, increasing the feature fusion capability of the distant target, and improving the BiFPN structure as shown in FIG. Figure 7b shown.

[0120] In the feature fusion process, different input features have different resolutions and different contributions to the output features. There are many multi-scale targets collected under natural conditions. BiFPN uses a fast normalization fusion module to balance the weights of different features, mine the deep information of the target, and reduce false detections and missed detections caused by complex environments. The relationship between its input and output is:

[0121]

[0122] In the formula, ω i The input feature I i The corresponding learning weights are ensured by applying ReLU in the later stage i ≥0), to avoid numerical instability, the initial learning rate ε=0.0001 is set, and the value of the normalized weight decreases between 0 and 1;

[0123] Taking feature layer P4 as an example, the intermediate features of P4 and output features The formulas are:

[0124]

[0125] In the formula, is the i-th level input feature, and Resize is the upsampling or downsampling operation;

[0126] In order to further improve efficiency, depthwise separable convolution is used in the feature fusion stage, and batch normalization and activation operations are added after convolution. The feature fusion network of the present invention is replaced with BiFPN, which can assign different weights to features of different scales, and perform cross-scale weight suppression or feature expression to enhance feature fusion, further improving target detection performance.

[0127] The third improvement is introduced as follows:

[0128] YOLOv8 network predicts the border coordinate loss L CloU The CloU (complete intersection over union) Loss is used, and its calculation is as follows:

[0129]

[0130] Where, L 10U Expression Figure 8 As shown ( Figure 8 In the above equation, w, h, (x, y) represent the width, height and center coordinates of the prediction box respectively; w gt 、h gt 、(x gt ,y gt ) represent the width, height and center coordinates of the real box respectively; W i , H i Respectively represent the width and height of the intersection; W g , H gRepresent the minimum frame width and height respectively), α is the weight function used to balance the parameters, v is the aspect ratio metric function used to measure the aspect ratio consistency. Although the CIoU loss function takes into account the overlapping area, center distance, and aspect ratio, in the prediction frame regression process, when the aspect ratio of the predicted frame and the true frame is linearly related, the penalty term of CIoU degenerates to 0. Therefore, both high-quality and low-quality anchor frames will be harmful to the regression loss.

[0131] The present invention selects the dynamic non-monotonic focusing mechanism WIoU to replace the original loss function, and its loss function is shown as follows:

[0132] L WIoU =rR WIoU L IoU ,R WIoU ∈[1,e),L IoU ∈[0,1]

[0133] Where, the distance focusing mechanism R WIoU L is used to enlarge the normal anchor box 10U , the non-monotonic focusing coefficient r is used to focus on anchor boxes of normal quality. WIoU Definition,r is defined as follows:

[0134]

[0135] In the formula, outlier in, Represents a dynamic sliding average, which allocates a smaller gradient gain when the value is large or small, reducing the impact on bounding box regression. α and δ represent hyperparameters. r dynamically gives the bounding box gradient gain by reducing the contribution of high-quality samples to the loss value, and reduces the harmful gradients generated by low-quality anchor boxes in the later stage of training, focusing on anchor boxes of ordinary quality, and improving the model's positioning ability.

[0136] WIoU removes the aspect ratio penalty term in CIoU, balances the impact of high-quality and low-quality anchor boxes on model regression, enhances the generalization ability of the model, and improves the overall performance of the model. Therefore, the present invention uses WIoU to replace the boundary loss function of the original model to optimize the model loss function.

[0137] The hardware configuration of the training model platform of the present invention is I99900K CPU, NvidiaA4000GPU, 8GB graphics card, and the software environment is win10 system, 64-bit, Python 3.9 version, Pytorch 2.0.0 framework, Cuda 11.7 version. The training parameters are shown in Table 1.

[0138] Table 1 Training parameters

[0139]

[0140]

[0141] The present invention provides an embodiment, which compares the improved YOLOv8-PBi model based on YOLOv8 with mainstream target detection network models such as SSD, RCNN, YOLOv5s, YOLOv5m, and YOLOv7-tiny. The experimental results are shown in Table 2.

[0142] Table 2 Comparison of detection results of different models

[0143]

[0144] As shown in Table 1, the method proposed in this application has better detection results.

[0145] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0146] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A roadside smart parking target detection method based on an improved YO10v8 model, characterized in that: include: Collect roadside parking images, preprocess and annotate them, and generate roadside parking datasets; Improve the traditional YOLOv8 model and build a roadside smart parking target detection model based on the improved YOLOv8 model; The roadside smart parking target detection model is trained based on the roadside parking dataset to obtain the final roadside smart parking target detection model; The image to be detected is input into the final roadside intelligent parking target detection model to obtain the detection result.

2. The roadside intelligent parking target detection method based on the improved YO10v8 model according to claim 1 is characterized in that: Collect roadside parking images, preprocess and annotate them, specifically: Collect a large number of roadside parking images at urban roadside parking spaces based on image acquisition equipment; Perform data enhancement processing on the collected roadside parking images; Annotate the enhanced roadside parking images, annotate the bounding box of each vehicle, the bounding box of the parking space, the license plate number of each vehicle, and the parking space number of the parking space; Generate a roadside parking dataset based on the processed images.

3. The roadside intelligent parking target detection method based on the improved YO10v8 model according to claim 2 is characterized in that: The collected roadside parking images are subjected to data enhancement processing, specifically: Data enhancement processing is performed on the collected roadside parking images based on the DM-GAN network model; Perform regional image enhancement on the roadside parking image after data enhancement.

4. The roadside intelligent parking target detection method based on the improved YO10v8 model according to claim 3 is characterized in that: The collected roadside parking images are enhanced based on the DM-GAN network model, specifically: Construct DM-GAN network model; The collected roadside parking images are enhanced based on the DM-GAN network model.

5. The roadside intelligent parking target detection method based on the improved YO10v8 model according to claim 4 is characterized in that: Construct the DM-GAN network model, specifically: Construct the DM-GAN network model, including the generator network module, the discriminator module and the loss function module; The generator network module is based on the U-Net network as the backbone, and the DC dual-channel attention mechanism is integrated into the jump connection part of the U-Net network. The MSDR multi-scale dense residual block is introduced at the bottom of the generator network module. The MSDR multi-scale dense residual block is designed by combining multi-scale feature extraction and dense residual. The discriminator module consists of the Patch-GAN network architecture; The loss function module includes perceptual loss, adversarial loss and SF loss.

6. The roadside smart parking target detection method based on the improved YO10v8 model according to claim 3 is characterized in that: The roadside parking image after data enhancement is subjected to regional image enhancement processing, specifically: Obtain a roadside parking image after data enhancement processing, which is an RGB format image; The roadside parking image after data enhancement is converted from RGB format to LAB format, and an L component image is obtained; Calculate the pixel mean of all pixels on the L component image to obtain the L component mean; Perform Gaussian smoothing on the L component image to obtain a Gaussian L component image; The pixel difference between each pixel point on the Gaussian L component image and the L component mean is calculated in sequence, and the salient area image is obtained according to the size of the pixel difference corresponding to each pixel point; Normalizing the salient region image to obtain a normalized image; Binarize the normalized image to obtain a binary image; Extract edge information of the binary image to obtain an edge image; The binary image and edge image are used as input, and the connected region detection method is used to extract the connected region of the normalized image, and the connected region is the region to be enhanced; An image enhancement algorithm is used to enhance all pixels in the enhanced area to obtain an image after regional image enhancement.

7. The roadside intelligent parking target detection method based on the improved YO10v8 model according to claim 1 is characterized in that: Improvements to the traditional YOLOv8 model are as follows: Faster Block is designed based on PConv in FasterNet. In the feature extraction C2f module of the traditional YOLOv8 model, Faster Block replaces Bottleneck as the main gradient flow branch by class inheritance. Replace the feature fusion network of the traditional YOLOv8 model with the BiFPN network structure; The dynamic non-monotonic focusing mechanism WIoU loss function is selected to replace the loss function of the traditional YOLOv8 model.

Images