Dense pedestrian target detection method based on improved YOLOv11

Abstract

The application discloses a dense pedestrian target detection method based on an improved YOLOv11, and realizes efficient detection of dense pedestrians by constructing an improved YOLOv11 model. The method first collects a dense pedestrian dataset and performs image preprocessing, and then constructs the improved YOLOv11 model. The specific improvement steps are as follows: the GhostConv module is used to replace the ordinary convolution in the backbone network, the redundant information in the feature map is reduced, and the calculation cost is reduced, so that the model is lightweight; the C3K2 module is replaced by C3K2-WTConv, the convolution receptive field is expanded, the low-frequency information in the image is effectively captured, and the dense target detection effect is significantly improved; the lightweight MLCA attention mechanism is introduced into the C2PSA module to form a new C2PSA-MLCA module, and the local and global information are combined to enhance the network expression ability. After the model construction is completed, training, testing and performance evaluation are performed. The application realizes model lightweight, improves the accuracy of target detection, and is suitable for traffic monitoring and safety monitoring in public places.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of target detection, and in particular to a dense pedestrian target detection method based on improved YOLOv11. BACKGROUND

[0002] Pedestrian detection, as one of the research directions in the field of target detection, has wide application value in intelligent transportation and security fields. This technology provides road environment perception capability for assisted driving systems through real-time identification and positioning of pedestrian targets, realizes accurate target tracking in road monitoring fields, and plays an important role in early warning and protection systems. Especially in crowded public places such as commercial complexes and tourist attractions, the pedestrian detection technology can provide reliable data support for passenger flow statistics, behavior analysis and safety management, and effectively improve the overall efficiency of intelligent security systems. However, the existing pedestrian detection algorithm still faces significant technical bottlenecks when dealing with large-scale dense scenes. Especially in the complex environment of high-density crowds and mutual occlusion, the algorithm often has difficulty in accurately identifying partially or completely occluded pedestrian targets, resulting in a significant decrease in detection accuracy. This technical problem has become a key factor restricting the practical application effect of pedestrian detection, and is also a core challenge that the current academic and industrial circles are concerned about. How to improve the robustness of the algorithm in crowded scenes and reduce the missed detection rate caused by occlusion is still a research focus that needs to be broken through.

[0003] In recent years, the rapid evolution of deep learning technology has driven the innovation of target detection methods. Under the framework of deep neural networks, detection algorithms have made breakthroughs mainly in two directions: one is the two-stage detection method represented by R-CNN and Fast R-CNN, which realizes target positioning through region proposal and feature extraction; the other is the single-stage detection method represented by YOLO series and SSD series, which directly predicts target position and category in an end-to-end manner. Compared with traditional detection methods, these new algorithms based on deep learning have shown significant advantages in detection accuracy and efficiency, and have become the mainstream technology route in the field of target detection.

[0004] In order to improve the target detection accuracy of the model and realize lightweight, the present application improves the YOLOv11 model. By introducing GhostConv convolution, the model is lightweight, thereby better adapting to the deployment requirements of terminal devices such as unmanned aerial vehicles. In addition, combined with WTConv convolution and MLCA attention mechanism, the detection accuracy of the model is effectively improved without significantly increasing the parameter amount, and the problem of dense small target detection is also effectively solved. SUMMARY

[0005] The purpose of the present application is to propose an improved YOLOv11-based dense pedestrian target detection method to solve the problems of low detection accuracy, large model parameter quantity and high computational complexity of dense small targets.

[0006] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0007] An improved YOLOv11-based dense pedestrian target detection method, comprising the following steps:

[0008] Step 1: Collect dense pedestrian dataset and perform image augmentation and label preprocessing on the dataset;

[0009] Step 2: Construct an improved YOLOv11-based dense pedestrian target detection model;

[0010] Among them, the step 2 specifically includes the following steps:

[0011] Step 2.1: In the backbone network of YOLOv11, replace the Conv ordinary convolution of the network with the GhostConv module;

[0012] Step 2.2: Introduce the WTConv module into the C3K2 module to form the C3K2-WTConv module, and replace the C3K2 module with the C3K2-WTConv module in the backbone network of YOLOv11;

[0013] Step 2.3: Introduce the MLCA attention mechanism into the C2PSA module to form the C2PSA-MLCA module, and replace the C2PSA module with the C2PSA-MLCA module in the backbone network of YOLOv11;

[0014] Step 3: Train the improved YOLOv11-based dense pedestrian target detection model;

[0015] Step 4: Perform performance evaluation on the improved YOLOv11-based dense pedestrian target detection model on the dense pedestrian test set.

[0016] 2. The improved YOLOv11-based dense pedestrian target detection method according to claim 1, wherein step 1 specifically comprises the following steps:

[0017] Step 1.1: Screen the collected dense pedestrian dataset, and manually delete the animation pictures and pictures with fewer pedestrians;

[0018] Step 1.2: Write a corresponding classification program, remove the ignore region and crowd annotations in the data set annotation file, keep the pedestrian, cyclist and partially visible person annotations and classify them as the same annotation Person, and convert the annotation file to YOLO format;

[0019] Step 1.3: Divide the data set into training set and test set;

[0020] 3. The improved YOLOv11-based dense pedestrian target detection method according to claim 1, wherein step 2.1 specifically comprises the following steps:

[0021] Step 2.1.1: First, use 1x1 ordinary convolution, batch normalization layer and a ReLu activation function to compress the channel number of the input picture, generate inherent feature mapping, and then apply a series of low-cost linear operations to the feature map Obtain more different feature maps to increase the features. The low-cost linear operation is composed of depth separable convolution, batch normalization layer and a ReLu activation function;

[0022] Step 2.1.2: The feature maps obtained by the low-cost linear operation and the feature maps obtained by the ordinary convolution, batch normalization and a ReLu activation function in step 2.1.1 are combined through Concat operation, and the final obtained feature map is the output.

[0023] 4. The improved YOLOv11-based dense pedestrian target detection method according to claim 1, wherein step 2.2 specifically comprises the following steps:

[0024] Step 2.2.1: WTConv first uses two-dimensional Haar wavelet transform to perform multi-level decomposition on the input image. The Haar wavelet transform uses four filters to decompose the image into four subbands: low-frequency component, horizontal high-frequency component, vertical high-frequency component and diagonal high-frequency component. In each level of wavelet transform, the image is down-sampled, so that the spatial resolution is halved, but the frequency information is more finely decomposed, thereby obtaining frequency components at different scales;

[0025] Step 2.2.2: Secondly, perform grouped convolution operation, use a small size of depth convolution kernel on each frequency subband, use a small convolution kernel of 7x7, and perform convolution operation on each subband after decomposition, the process is as follows:

[0026] [X LL ,X LH ,X HL ,X HH ]=Conv([f LL ,f LH ,fHL ,f HH ],X)

[0027] where X is the input tensor, f LL is a low-pass filter, f LH , f HL , f HH is a set of high-pass filters, X LL is the low-frequency component, X LH , X HL , X HH are the high-frequency components of horizontal, vertical and diagonal lines respectively;

[0028] Step 2.2.3: IWT operation is linear, which can losslessly reconstruct the WT convolution result and the grouped convolution result to the original space. After the grouped convolution operation is completed, the convolution result of each sub-band is reconstructed into a complete output using the inverse wavelet transform of IWT, the process is as follows:

[0029] Y = IWT(Conv(W, WT(X)))

[0030] where X is the input tensor, W is a weight tensor of k x k depth convolution kernel, and the number of input channels is four times that of X.

[0031] 5. The improved YOLOv11-based dense pedestrian target detection method according to claim 1, wherein step 2.3 specifically comprises the following steps:

[0032] Step 2.3.1: The input feature map is first processed by local average pooling and global average pooling, and secondly, the features after local pooling and global pooling are both subjected to a 1D convolution for feature conversion, so that the features are rearranged and adapted to subsequent operations.

[0033] Step 2.3.2: After 1D convolution, the features after local pooling are rearranged. This process is equivalent to a kind of feature selection, which strengthens the attention to useful features.

[0034] Step 2.3.3: After 1D convolution, rearrangement and unpooling operation (UNAP) on the features after global pooling, the features are combined with the local pooling features through addition operation. This process fuses global context information in the feature map.

[0035] Step 2.3.4: After the combined features after local and global attention processing, the features are subjected to unpooling operation (UNAP) again, and then combined with the original input features through multiplication operation, restoring to the original spatial dimension.

[0036] 6. The improved YOLOv11-based dense pedestrian target detection method according to claim 1, wherein step 3 specifically comprises the following steps:

[0037] Step 3.1: dividing the preprocessed dense pedestrian data set into a training set and a test set according to a ratio of 7:3;

[0038] Step 3.2: setting the number of iterations to 200, setting the batch size batch_size to 16, and setting the optimizer optimizer to auto;

[0039] Step 3.3: training the improved YOLOv11-based dense pedestrian target detection model according to the set parameters to obtain the trained improved YOLOv11-based dense pedestrian target detection model.

[0040] The beneficial effects of the present application are as follows:

[0041] 1. In the backbone network of YOLOv11, the present application replaces the Conv ordinary convolution with the GhostConv module, which significantly reduces the calculation cost of redundant information while retaining the feature expression ability by separating the generation process of the feature map, so that the model can be deployed on mobile terminals.

[0042] 2. In the backbone network of YOLOv11, the present application replaces the C3K2 module with the C3K2-WTConv module, and wavelet transform effectively captures more context information in the image without increasing the calculation amount, thereby improving the accuracy of target positioning. And WTConv effectively alleviates the problem of fine-grained feature loss caused by the excessive response of traditional convolution to high-frequency information by enhancing low-frequency feature extraction, thereby improving the accuracy of target detection.

[0043] 3. In the backbone network of YOLOv11, the present application introduces MLCA attention mechanism into C2PSA to form C2PSA-MCLA module, which fuses multi-dimensional information through local and global pooling methods, enhances the feature expression ability of pedestrians, and through 1D convolution operation, it significantly reduces the parameter complexity and calculation amount while improving the accuracy of dense pedestrian target detection. BRIEF DESCRIPTION OF DRAWINGS

[0044] The specific embodiments and the accompanying drawings will be described in detail below.

[0045] Fig. 1 Figure 1 is a schematic diagram of the overall structure of the improved YOLOv11 network;

[0046] Fig. 2 Figure 2 is a schematic diagram of the structure of GhostConv.

[0047] Fig. 3 A structural schematic diagram of WTConv;

[0048] Fig. 4 A structural schematic diagram of an MLCA attention mechanism;

[0049] Fig. 5 A flowchart of the dense pedestrian target detection method based on the improved YOLOv11 proposed in the present application. DETAILED DESCRIPTION

[0050] In order to make the embodiments of the present application and the prior art more clear and easy to understand, the present application will be described in detail below with reference to the accompanying drawings.

[0051] Referring to Fig. 5 , the steps of obtaining the dense pedestrian target detection model based on the improved YOLOv11 are as follows:

[0052] Step 1: Collect dense pedestrian datasets using various large search engines and perform image augmentation and label preprocessing on the datasets;

[0053] Step 1.1: In the collected dataset images for augmentation, there are many images that are not suitable for training, such as cartoon animation images, images with too few pedestrians in the crowd, and images with too high blur, etc. Therefore, the collected dense pedestrian dataset is manually screened, and the images unsuitable for training are deleted, then the pedestrians in each image are labeled using the rectangular labeling in the LabelImg image labeling tool, and the labels are uniformly set as the person class, and after the labeling is completed, the corresponding txt label file is obtained, recording the information of the bounding box;

[0054] Step 1.2: In the image labels of part of the collected dataset, there are 5 types of labels: pedestrians, riders, partially visible persons, ignored regions, and crowds, so a corresponding classification program is written to remove the ignored region and crowd labels in the annotation file in the dataset, and keep the pedestrian, rider, and partially visible person labels and classify them as the same label Person, and convert the annotation file into YOLO format;

[0055] Step 1.3: Integrate the collected dataset, images, and labels, and divide them into training set and test set according to the ratio of 7:3;

[0056] Step 2: Build a dense pedestrian target detection model based on the improved YOLOv11;

[0057] In order to solve the problem of dense pedestrian target detection, the YOLOv11 network is improved from the following three aspects, as shown in Figs. 1 to 4 ​

[0058] Step 2.1: In order to reduce the computational load and complexity of the network, in the backbone network of YOLOv11, the Conv ordinary convolution of the 0th, 1st, 3rd, 5th, 7th layers of the network is replaced by the GhostConv module. First, 1x1 convolution is performed on the input feature map to generate intrinsic feature maps with fewer channels, which aggregate the information features between channels. Then, depthwise convolution is applied to the output feature map of 1x1 convolution to generate redundant feature maps (Ghost feature maps) with the same number of initial feature maps. Finally, the intrinsic feature maps and the redundant feature maps are concatenated in the channel dimension to output the final feature maps.

[0059] Step 2.2: Introduce the WTConv module into the C3K2 module to form the C3K2-WTConv module, and replace the C3K2 module of the 2nd, 4th, 6th, 8th layers of the backbone network of YOLOv11 with the C3K2-WTConv module.

[0060] Step 2.3: Introduce the MLCA attention mechanism into the C2PSA module to form the C2PSA-MLCA module, and replace the C2PSA module of the 10th layer of the backbone network of YOLOv11 with the C2PSA-MLCA module.

[0061] Step 2.1 includes the following steps:

[0062] Step 2.1.1: First, use 1x1 ordinary convolution, batch normalization layer and a ReLu activation function to compress the channel number of the input picture to generate intrinsic feature maps, and then apply a series of low-cost linear operations to the feature maps to obtain more different feature maps and increase redundant features. If the original convolution needs to generate a feature map with C channels, the convolution will generate a feature map with C / 2 channels, and then generate a new C / 2 new channels through 3x3 depthwise convolution. The low-cost linear operation is composed of depthwise convolution, batch normalization layer and a ReLu activation function;

[0063] Step 2.1.2: The feature maps obtained by low-cost linear operations and the feature maps obtained by 1x1 ordinary convolution, batch normalization and a ReLu activation function in step 2.1.1 are concatenated through Concat operation. At this time, the channel number of the output feature map is restored to the level of the original convolution, and the final obtained feature map is the output.

[0064] Step 2.2 includes the following steps:

[0065] Step 2.2.1: WTConv first uses a two-dimensional Haar wavelet transform to perform multi-level decomposition on the input image. Among them, the Haar wavelet transform uses four filters to decompose the image into four subbands: low-frequency component, horizontal high-frequency component, vertical high-frequency component, and diagonal high-frequency component. In each level of wavelet transform, the image is down-sampled, so that the spatial resolution is halved, but the frequency information is more finely decomposed, thereby obtaining frequency components at different scales;

[0066] Step 2.2.2: Secondly, a convolution operation is performed, using a small size of deep convolution kernel on each frequency subband, using a 7x7 small convolution kernel, the decomposed subbands are convolved, the process is as follows:

[0067] [X LL ,X LH ,X HL ,X HH ]=Conv([f LL ,f LH ,f HL ,f HH ],X)

[0068] Where X is the input tensor, f LL is a low-pass filter, f LH , f HL , f HH is a set of high-pass filters, X LL is the low-frequency component, X LH , X HL , X HH are the horizontal, vertical and diagonal high-frequency components respectively;

[0069] Step 2.2.3: IWT operation is linear, which can losslessly reconstruct the convolution result to the original space, after the convolution operation is completed, the IWT inverse wavelet transform is used to combine the convolution results of each subband into a complete output, the process is as follows:

[0070] Y=IWT(Conv(W,WT(X)))

[0071] Where X is the input tensor, W is a weight tensor of a k x k deep convolution kernel, and the number of input channels is four times that of X.

[0072] Wherein, step 2.3 includes the following steps:

[0073] Step 2.3.1: input feature map X∈R B×C×H×WFirst, the feature map is compressed to a fixed size by adaptive average pooling (AdaptiveAvgPool2d) to extract local spatial information, reduce the size of the feature map, and reduce the computational complexity of subsequent calculations. The process is as follows:

[0074] X local = AdaptiveAvgPool2d(X, S)

[0075] where B is the batch size, C is the number of input feature map channels, H and W are the height and width of the input feature map, respectively;

[0076] Step 2.3.2: Then, local feature extraction is performed: the pooled local feature X local ∈R B×C×S×S is subjected to 1D convolution to calculate the local dependence between channels, where B is the batch size, C is the number of input feature map channels, and S is the local block size;

[0077] Step 2.3.3: The local feature is further subjected to global average pooling (GAP) to generate global channel feature X glocal ∈R B×C×1×1 , which is then subjected to 1D convolution to capture global channel weights. This process is equivalent to a kind of feature selection, which strengthens the attention to useful features;

[0078] Step 2.3.4: The global pooling feature weight is combined with the local pooling feature weight through an addition operation. This process integrates global context information in the feature map, as follows:

[0079] W final = λ·W local +(1-λ)·W glocal

[0080] where W loca = Conv1d(X local ) and W glocal = Conv1d(X glocal ) are the local weight and global weight, respectively;

[0081] Step 2.3.5: Finally, the feature map after local and global attention processing is combined, then subjected to an unpooling operation (UNAP), and then combined with the original input feature through a multiplication operation to restore the original spatial dimension, enhance important features, and suppress redundant information, as follows:

[0082]

[0083] Wherein, the reverse pooling operation adjusts the weight back to the original input size HxW through adaptive up sampling (AdaptiveAvgPool2d) to ensure alignment with the input feature map, and sigma is a sigmoid function. represents element-wise multiplication.

[0084] Step 3: training the dense pedestrian target detection model based on the improved YOLOv11 includes the following steps:

[0085] Step 3.1: divide the pre-processed dense pedestrian dataset into a training set and a test set according to a ratio of 7:3, and input the training set into the improved YOLOv11;

[0086] Step 3.2: design a training program, uniformly adjust the size of the picture to 640x640 before training, and perform data enhancement operations including random cropping, random flipping, etc. to increase the amount of data and improve the generalization ability of the model, set the number of iterations epochs to 300, the batch size batch_size to 16, and the optimizer optimizer to auto;

[0087] Step 3.3: train the dense pedestrian target detection model based on the improved YOLOv11 according to the set parameters, calculate the total loss function and optimize the model parameters with the data to obtain the trained dense pedestrian target detection model based on the improved YOLOv11;

[0088] Wherein, the loss function includes a positioning loss (bounding box regression loss), a confidence loss and a classification loss. The positioning loss is used to measure the difference between the predicted frame center coordinates (x, y) and size (w, h) and the real frame, and the calculation formula is as follows:

[0089]

[0090] Wherein, S 2 represents that the image is divided into SxS grid, B represents the number of candidate frames predicted by each grid, is an indicator function, which is 1 when the jth candidate frame of the ith grid is responsible for detecting objects, otherwise it is 0, and lambda coord is a weight parameter for enhancing the weight of the position loss, x i ,y i represents the offset of the predicted frame center relative to the grid, is the true value;

[0091] The confidence loss is used to measure the accuracy of the confidence of the predicted frame containing the object (i.e. the IoU of the predicted frame and the real frame), and the calculation formula is as follows:

[0092]

[0093] where C i represents the confidence of the prediction box (the value range is 0-1), represents the real confidence (1 if the candidate box has the maximum IoU with the real box, otherwise 0), and λ noobj represents a weight parameter, which is used to reduce the loss weight of the region without object, represents an indicator function, which is 1 when the grid does not contain an object.

[0094] The classification loss is used to measure the accuracy of the class prediction of the prediction box, and the calculation formula is as follows:

[0095]

[0096] where, represents an indicator function, which is 1 when the grid contains an object, and p i (c) represents the predicted class probability, represents the real class probability.

[0097] In summary, the total loss of YOLOv11 is the weighted sum of the above three parts, and the calculation formula is as follows:

[0098] L total = L coord + L conf + L class

[0099] Step 4: After training, a model saved every 10 epochs will be obtained, so compare the total loss function and accuracy of each 10 epochs in the training process, select the model with the best performance, and perform performance evaluation on the dense pedestrian test set using the model.

[0100] where the performance evaluation of the present application includes three key indicators: precision (p), recall (R), and mean average precision (mAP), and the specific formula is as follows:

[0101]

[0102]

[0103]

[0104] where TP (True Positives) is the number of correctly predicted positive samples, FN (False Negatives) is the number of incorrectly predicted negative samples, and FP (False Positives) is the number of incorrectly predicted positive samples.

[0105] The improved YOLOv11-based dense pedestrian target detection method provided by the application fully pays attention to the redundant information in the dense pedestrian feature map, enhances the feature recognition ability of the model to the shape rather than the texture, enables the model to improve the accuracy of target detection while realizing lightweight, and is particularly suitable for public security fields such as subway station passenger flow analysis and urban crossroad pedestrian detection.

[0106] The above is only the preferred embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art should be within the protection scope of the present application according to the technical solution and the inventive concept of the present application, and any equivalent replacement or change should be within the protection scope of the present application.

Claims

1. A dense pedestrian target detection method based on improved YOLOv11, characterized in that: The steps include: Step 1: Collect a dense pedestrian dataset and perform image preprocessing on the dataset; Wherein, the step 1 specifically includes the following steps: Step 1.1: Filter the collected dense pedestrian dataset and manually delete animated images and images with a small number of pedestrians; Step 1.2: Write a corresponding classification program to remove the ignored area and crowd labels from the annotation files in the dataset, retain the pedestrian, rider, and partially visible person labels, and classify them as the same label Person. Then convert the annotation files into YOLO format. Step 1.3: Divide the dataset into training and test sets; Step 2: Build a dense pedestrian target detection model based on improved YOLOv11; Wherein, the step 2 specifically includes the following steps: Step 2.1: In the backbone network of YOLOv11, replace the ordinary convolution of the network with the GhostConv module; Step 2.2: Introduce the WTConv module into the C3K2 module to form the C3K2-WTConv module, and replace the C3K2 module with the C3K2-WTConv module in the backbone network of YOLOv11; Step 2.3: Introduce the MLCA attention mechanism into the C2PSA module to form the C2PSA-MLCA module, and replace the C2PSA module with the C2PSA-MLCA module in the YOLOv11 backbone network; Step 3: Train the dense pedestrian target detection model based on the improved YOLOv11; Wherein, the step 3 specifically includes the following steps: Step 3.1: Divide the preprocessed dense pedestrian dataset into a training set and a test set in a ratio of 7:3; Step 3.2: Set the number of iterations to 200, the batch size to 16, and the optimizer to auto. Step 3.3: Train the improved YOLOv11-based dense pedestrian target detection model according to the set parameters to obtain a trained improved YOLOv11-based dense pedestrian target detection model; Step 4: Evaluate the performance of the trained dense pedestrian target detection model based on the improved YOLOv11 on the dense pedestrian test set.

2. The dense pedestrian target detection method based on improved YOLOv11 according to claim 1 is characterized in that: Step 2.1 specifically includes the following steps: Step 2.1.1: First, use a 1×1 ordinary convolution, a batch normalization layer and a ReLu activation function to compress the number of channels of the input image to generate an inherent feature map, and then apply a series of low-cost linear operations to the feature map. Obtain more different feature maps and increase features, among which low-cost linear operations It consists of a depth-wise separable convolution, a batch normalization layer, and a ReLU activation function; Step 2.1.2: Combine the feature map obtained through the low-cost linear operation and the feature map obtained through ordinary convolution, batch normalization and a ReLU activation function in step 2.1.1 through the Concat operation. The final feature map is the output.

3. The dense pedestrian target detection method based on improved YOLOv11 according to claim 1 is characterized in that Step 2.2 specifically includes the following steps: Step 2.2.1: WTConv first uses a two-dimensional Haar wavelet transform to perform a multi-level decomposition of the input image. The Haar wavelet transform uses four filters to decompose the image into four subbands: low-frequency component, horizontal high-frequency component, vertical high-frequency component, and diagonal high-frequency component. In each level of the wavelet transform, the image is downsampled, which halves the spatial resolution, but the frequency information is decomposed more finely, thereby obtaining frequency components at different scales. Step 2.2.2: Next, perform a convolution operation. Use a small-sized depth convolution kernel on each frequency subband. Use a small 7×7 convolution kernel to perform a convolution operation on each decomposed subband. The process is as follows: [X LL ,X LH ,X HL ,X HH ]=Conv([f LL ,f LH ,f HL ,f HH ],X) Where X is the input tensor, f LL is a low-pass filter, f LH ,f HL , f HH is a set of high-pass filters, X LL is the low-frequency component, X LH , X HL , X HH are the high-frequency components of horizontal, vertical, and diagonal lines, respectively; Step 2.2.3: The IWT operation is linear and can losslessly reconstruct the convolution result into the original space. After the convolution operation is completed, the IWT inverse wavelet transform is used to re-synthesize the convolution results of each sub-band into a complete output. The process is as follows: Y = IWT(Conv(W, WT(X))) Where X is the input tensor and W is the weight tensor of a k×k depthwise convolution kernel with four times the number of input channels as X.

4. The dense pedestrian target detection method based on improved YOLOv11 according to claim 1 is characterized in that Step 2.3 specifically includes the following steps: Step 2.3.1: The input feature map is first processed by local average pooling and global average pooling. Secondly, the local pooled features and the global pooled features are transformed by a 1D convolution so that the features are rearranged and adapted to subsequent operations. Step 2.3.2: After local pooling, use 1D convolution and then rearrange it. This process is equivalent to feature selection, which strengthens the focus on useful features. Step 2.3.3: After 1D convolution, rearrangement and unpooling operations, the global pooled features are combined with the local pooled features through addition operations. This process integrates global context information into the feature map. Step 2.3.4: Finally, after the feature maps processed by local and global attention are combined, they are again unpooled and then combined with the original input features through multiplication to restore the original spatial dimensions.

Images