Deformation target detection method based on adaptive feature extraction network and attention mechanism

Abstract

The invention discloses a deformation target detection method based on an adaptive feature extraction network and an attention mechanism, and belongs to the technical field of image recognition. The method comprises the following steps: constructing a high-quality electric car elevator entering data set; a tramcar entering detection model is constructed; training a tramcar ladder entering detection model; and inputting a to-be-identified image into the trained model to carry out tramcar entering detection, and verifying a detection result. The problems of missing detection, false detection and the like due to feature deformation of a detection target when a monitoring view angle is fixed, a space is narrow and a large number of people exist in an existing detection model based on deep learning are solved. Based on the above problems, a tramcar entering data set is constructed, and a tramcar entering detection model is trained. According to the model, a residual structure in a backbone network C3 module is fused with deformable convolution, a SimAM attention mechanism is introduced into a backbone network, a weighted bidirectional feature pyramid network is introduced into a neck network, and the representation capability of detail features during target deformation is enhanced, so that the accuracy of target detection is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular, to a deformed object detection method based on an adaptive feature extraction network and an attention mechanism. Background Art

[0002] In the field of computer vision, deformed objects specifically refer to objects whose appearance features undergo non-rigid or geometric deformations due to changes in imaging conditions. Such objects widely exist in real-world scenarios, and their deformation features mainly include: geometric deformation, non-rigid deformation, occlusion deformation, etc. The detection of electric vehicles in an elevator scenario faces typical compound deformation problems. Due to the limited space in the elevator car, cameras are usually installed in a high-position upward shooting manner, resulting in the following characteristics in the imaging: the target shows obvious perspective compression in the vertical direction, local high-light occlusion caused by the reflection of the metal inner wall of the car, and mutual occlusion when multiple targets overlap. These factors together cause the electric vehicle target to present complex deformation features in the image.

[0003] Traditional methods mainly deal with deformed objects through geometric transformation and feature engineering. Among them, methods based on geometric transformation include algorithms such as TPS and STN, which perform deformation correction on the input image through mathematical transformations (such as rotation, scaling, and shearing) to restore the standard pose of the target; methods based on feature engineering mainly include scale-invariant feature transform, which constructs rotation-invariant features through key-point detection and histogram of oriented gradients. However, traditional methods rely on preset geometric priors, are difficult to adapt to complex non-rigid deformations, and have a large computational overhead. They may face challenges in computing and storage resources when dealing with large-scale datasets or high-resolution images. Therefore, the application of traditional object detection methods in actual scenarios has many limitations and cannot achieve effective and accurate detection work. With the development of deep learning, object detection algorithms centered on convolutional neural networks (CNNs) have made remarkable progress in general scenarios, but there are still bottlenecks in deformed object detection. When performing deformed object detection, the fixed-size RoIPooling operation in Faster R-CNN will destroy the local feature structure of the deformed object, and the generation of candidate regions depends on preset anchor boxes, making it difficult to match the shape of the target after perspective deformation; YOLO series algorithms regard object detection as a single regression problem and directly predict the bounding box and category on the feature map. However, when performing deformed object detection, the standard convolution operation uses a fixed sampling grid, making it difficult to capture the local features of non-rigid deformations (such as a distorted wheel). When occlusion causes deformation of an object, limited important features can be extracted through the feature extraction network. Therefore, existing detection methods have certain limitations in the deformed object detection task.

[0004] To address the above problems, the present invention takes the detection of deformed objects as the research task, uses the tram entering the elevator monitoring images as an example research object, and introduces the following improvements in the YOLOv5 framework: introducing a deformable convolution module, dynamically adjusting the sampling positions by predicting the offsets of the convolution kernels to adapt to the deformed features; integrating a weighted attention module, enhancing the response of key regions through feature re-weighting, and proposing a deformed object detection method based on an adaptive feature extraction network and an attention mechanism. Summary of the Invention

[0005] The main technical problem to be solved by the present invention is the low target detection accuracy caused by the lack of overall target features and incomplete local detail features during the target detection process of objects with non-rigid or geometric deformations in their appearance features due to changes in imaging conditions.

[0006] To address the above technical problems, the present invention proposes a deformed object detection method based on an adaptive feature extraction network and an attention mechanism. The method includes the following steps:

[0007] Step 1: Construct a tram entering the elevator dataset with rich types;

[0008] For the tram entering the elevator detection task based on machine vision, a high-quality dataset is crucial for accurately detecting target features. The performance of the model is often closely related to the quality of the dataset. The quantity, diversity, and annotation accuracy of the dataset can directly reflect the quality of the dataset. Since there is no publicly available tram entering the elevator dataset, to solve the current dataset vacancy problem, this method constructs a tram entering the elevator dataset with rich types. The image data in the dataset mainly comes from public datasets, web crawlers, elevator monitoring videos of a certain community, and manual shooting. The dataset consists of a total of 6,673 images. In addition to including images of different angle types, it also includes some images with exposure affected by lighting, and images with personnel and object occlusion.

[0009] First, this method uses the publicly available 51CTO pedestrian and vehicle dataset and the KITTI dataset. These two datasets contain a large amount of pedestrian and bicycle data. However, due to the relatively single scene and the lack of electric vehicle data, web crawlers are used to collect network images and take on-site pictures to expand the dataset, increasing the number of pictures in different fixed perspectives, local exposures, and occlusion scenarios to enrich the diversity of the data. In addition, considering that there are too few negative samples in the dataset and the detection is easily interfered by structurally similar bicycles, walkers, wheelchairs, etc., the existing negative sample data is randomly transformed: geometric transformation (rotation, flipping, scaling), color transformation (such as changing brightness, contrast, hue) to generate new negative sample data, effectively improving the balance between the accuracy and recall rate of the model. Finally, all the data is processed using image enhancement techniques, such as random cropping, flipping, blurring, adding noise, gamma transformation, etc., to obtain more image data, enrich the dataset, improve the generalization ability and robustness of the model. Further, the integrated and expanded dataset is normalized to unify information such as the file format and size of the images, and an annotation software is used to manually annotate the dataset images to provide accurate target location and category information for the dataset for model training and evaluation.

[0010] Step 2: Build a tram entering ladder detection model;

[0011] Design a detection model suitable for deformed targets. This detection model is designed based on the YOLOv5 model and consists of three parts: a backbone network, a neck network, and a prediction network.

[0012] The first stage: The main task of the backbone network in the detection model is to extract the features of the input tram entering ladder image in Step 1 and pass the features to the subsequent network layers for tram entering elevator detection. The backbone network consists of a Focus structure, a C3 structure, and an SPP structure. The Focus structure fuses the picture information onto the channels through slicing operations and then extracts the tram target features, making the feature extraction more sufficient. The C3 structure divides the input feature map into two sub-maps. One sub-map is sent to a convolutional layer composed of three consecutive 3×3 convolutional operations for feature extraction, while the other sub-map is directly concatenated with the result after convolution, effectively reducing the computational complexity of the classification problem and enhancing the learning ability of the backbone network. The SPP structure enhances the extraction effect of tram features through window pooling of different sizes.

[0013] In the actual tram entering elevator detection scenario, due to fixed perspective and limited space, the target deforms, resulting in relatively limited key tram detail feature information. Feature learning based on the global context is interfered by local deformation features, making it difficult to accurately obtain deeper feature information and resulting in weak key feature learning ability of the backbone network. By integrating the deformable convolution into the residual structure in the C3 structure of the backbone network and changing it into a CD2Bottleneck module, and integrating it into the backbone network to replace the original C3 structure, it can adapt to the changes in detail feature information caused by the appearance and form changes of different detected targets. Additionally, a weighted attention mechanism SimAM is added to improve the ability to extract target detail features, enhance the feature learning ability of the deep network, obtain a tram feature map with stronger semantic information, and enable the tram entering elevator detection model to accurately extract the key feature information of the detection target and accurately perform the detection task of deformed targets.

[0014] The second stage: The neck network uses the BiFPN structure improved from the PANet structure, mainly for fusing different tram feature layers extracted by the backbone network. The PANet structure consists of two types of feature pyramids. One is the top-down feature pyramid structure, which extracts multi-scale tram feature maps from the detection head through a series of convolutional and downsampling operations. The feature maps in the top-down path have high resolution and semantic information. The other is the bottom-up inverted pyramid structure, which extracts multi-scale tram feature maps from the backbone network through a series of convolutional and upsampling operations. These feature maps have different resolutions and semantic information. Then, the bottom-up and top-down feature maps are fused layer by layer through the concat operation, enabling features with different resolutions to complement and fuse with each other. The BiFPN structure applies two-way multi-scale feature fusion in the top-down and bottom-up directions, enhancing the fusion of tram feature information between the same network layers and the transmission of tram feature information between different layers, enabling the network to have a higher important feature purification efficiency, reducing the interference of the overall tram feature deformation on the target detection task, and improving the detection accuracy of tram targets with different deformations of the target size.

[0015] The third stage: The prediction network consists of convolutional layers, fully connected layers, and activation functions, mainly for feature extraction and spatial compression of the input tram feature map. The prediction network has a total of 3 decoupled heads, which are respectively used to predict target boxes of different scales. Each decoupled head contains a series of convolutional layers and fully connected layers, used to extract features from the input tram feature map and predict the target category and bounding box. These decoupled heads independently correspond to tram targets of different scales, enabling the detection model to better adapt to target detection tasks of different sizes and improving the classification and regression performance of the tram entering elevator detection model.

[0016] Step 3: Train the tram entering elevator detection model;

[0017] When training the network model, the SGDM optimizer is selected to implement the iterative update of network parameters. The batch_size of network training is set to 32, the initial learning rate is set to 0.001, the momentum is 0.937, the epoch is set to 100, and the input size is 640×640. When training the tram entering elevator detection model, a balance can be found between training speed and accuracy by introducing an adaptive learning rate and an early stopping mechanism. The adaptive learning rate can dynamically adjust the learning rate according to the change of loss to accelerate the convergence speed and improve the accuracy of the model. The early stopping mechanism can avoid overfitting and stop training in advance when the validation loss no longer decreases, thus saving time and computing resources.

[0018] Step 4: Input the image to be recognized into the trained tram entering elevator model for detection to verify the detection effect.

[0019] The present invention has the following advantages compared with the prior art:

[0020] 1. A tram entering elevator dataset with a large amount of data, rich diversity, and accurate annotation is constructed, including a total of 6673 image samples. According to different image types, it includes 1561 images with special perspectives (producing distortion), 1081 locally exposed images, and 2334 samples with missing features. The unified size is 640×640 pixels, and the image format is JPEG. The created dataset solves the problems of single scene and lack of electric vehicle data volume in the existing dataset, providing a reliable data source for training the tram entering elevator detection model.

[0021] 2. A tram entering elevator detection model improved based on the YOLOv5 model is designed. Deformable convolution is fused in the backbone network and a weighted attention mechanism is added, and the BiFPN structure is adopted in the neck network. This model can adapt to the appearance and morphological changes of different targets during detection, adaptively adjust the attention of the model, better capture detailed features and handle deformations, thereby improving the accuracy and stability of deformed target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a schematic flow chart of the deformed target detection method based on the adaptive feature extraction network and the attention mechanism provided by the present invention.

[0023] Figure 2 is a schematic diagram of the network structure of the tram entering elevator detection model of the present invention.

[0024] Figure 3 is a schematic diagram of the Focus module of the tram entering elevator detection network model of the present invention.

[0025] Figure 4Schematic diagram of the CD2 module of the tram entering ladder detection network model of the present invention.

[0026] Figure 5 Schematic diagram of the DCNv2 structure of the tram entering ladder detection network model of the present invention.

[0027] Figure 6 Schematic diagram of the SimAM module of the attention mechanism of the tram entering ladder detection model of the present invention.

[0028] Figure 7 Schematic diagram of the BiFPN structure of the neck network of the tram entering ladder detection model of the present invention. Specific implementation mode

[0029] The present invention mainly realizes a deformation target detection method based on an adaptive feature extraction network and an attention mechanism. The specific method adopted by the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Specifically, the flow schematic diagram of the deformation target detection method based on the adaptive feature extraction network and the attention mechanism is shown in Appendix Figure 1 as shown, and it includes the following 5 steps: S1: Construct a tram entering ladder detection data set;

[0031] S2: Construct a tram entering ladder detection model; S3: Train the tram entering ladder detection model; S4: Input the image to be detected into the trained tram entering ladder detection model for testing; S5: The tram entering ladder detection model automatically detects the input image.

[0032] (1) For S1: Construct a high-quality tram entering ladder detection data set

[0033] The quality of the data set is crucial for the performance of the model. The data set suitable for tram entering ladder detection needs to include the deformation target problems caused by fixed perspective, narrow space, and occlusion, which the model mainly solves. Currently, the existing public data sets, the 51CTO pedestrian and vehicle data set and the KITTI data set, have the problems of relatively single shooting perspectives and shooting scenes, and at the same time, the amount of electric vehicle data is small. Therefore, according to the current deformation target detection problem to be solved, part of the pictures are collected through web crawling. In addition, the actual scene application problem also needs to be considered. According to the position and angle of the actual scene monitoring camera, sufficient electric vehicle data and other vehicle data are fully collected. Finally, the integrated images are used with data augmentation and data normalization algorithms for data expansion and cleaning, and the data set is manually labeled using the Labelme annotation software.

[0034] (2) For S2: Construct a tram entering ladder detection model improved based on YOLOv5s

[0035] The tram entering the elevator detection model mainly consists of three parts: the backbone network, the neck network, and the prediction network. The network structure of the model is as shown in Figure 2 Figure []. In the first stage, the backbone network is mainly responsible for feature extraction of the input image to be detected. The specific steps are as follows:

[0036] The Focus slicing operation splits the input feature map into four smaller feature maps, performs convolution operations on each small feature map, and finally stacks these small feature maps together to form an output feature map. Specifically, as shown in Figure 3 Figure [], for an image with an input size of 640×640×3, by sampling every other pixel, four independent feature layers are obtained, and the size of each feature layer is 320×320×3. Subsequently, these four independent feature layers are stacked together to form a feature map with a size of 320×320×12, where the width and height information is integrated into the channel information, and the number of input channels is quadrupled. Compared with the original 3-channel feature map, the feature map spliced together in this way has 12 channels.

[0037] The Conv operation mainly consists of three parts: convolution, Batch Normalization, and the SiLu activation layer. Among them, Batch Normalization can prevent overfitting and accelerate convergence, improving the stability of the model; the SiLu activation function is a weighted linear combination of Sigmoid.

[0038] The C2D operation mainly includes two parts. The first part uses the above-mentioned specified multiple Bottlenecks stacked, and the second part only passes through a basic convolution module. Finally, the two branches are concatenated. Specifically, as shown in Figure 4 Figure []. Among them, DCNv2 is introduced in the Bottleneck module. DCNv2 introduces a deformable convolutional layer, making the sampling position no longer limited to the traditional rectangular shape, but can adaptively adjust the shape of the convolutional kernel and the sampling position, with a stronger receptive field adjustment ability, and can flexibly adapt to complex target deformations. As shown in Figure 5 Figure [], the construction of the DCNv2 algorithm includes two key steps. First is the offset generation stage, where the sampling point offsets of the convolutional kernel on the input feature map are generated through convolution operations, covering the x and y directions. Second is the sampling and convolution stage, where bilinear interpolation is used with the input feature map and the calculated offsets to accurately determine the sampling point positions of the convolutional kernel on the input feature map. Through these sampling points, convolution operations are finally performed to process the feature map.

[0039] Introducing DCNv2 does not increase the number of model parameters. However, in practical applications, the frequent use of deformable convolutions in DCNv2 significantly increases the computational time. To reduce the model's computational load and maintain its light weight, only the Bottleneck residual structure in the original backbone network module is fused with deformable convolutions, changing it to the CD2Bottleneck module.

[0040] The SimAM attention mechanism added to the structure is a 3D attention mechanism that can directly calculate three-dimensional attention weights, fully utilize the information of three-dimensional data, and enable the model to better capture the detailed features of the target. Its overall structure is as Figure 6 shown. The SimAM attention mechanism is based on neuroscience theory. Utilizing the property that neurons with more information are more prominent compared to their adjacent neurons, when dealing with relevant visual tasks, neurons carrying key information will be assigned higher weights. It has a simple structure and does not require the introduction of additional parameters. It can directly derive three-dimensional weight values for the feature map, shortening the calculation time of attention weights and effectively improving the model's interference suppression ability and the accuracy of deformable object detection.

[0041] The SPP operation first halves the input channels through a standard convolution module, then uses three maximum pooling operations with a pooling kernel size of 5, connected in a serial and parallel manner. The results of the three maximum pooling operations are concatenated with the data without pooling operation. Finally, the number of channels after merging is twice the original. After convolution, batch normalization, and ReLu activation function operations, the operation is completed. Through the SPP operation, local features and global features can be fused, enriching the expressive ability of the feature map and significantly reducing the computational load while obtaining feature information at different scales.

[0042] In the second stage, the neck network mainly extracts features from the input images to be detected and fuses information from different feature map levels to improve object detection performance. The BiFPN network structure adopted by the neck network is as Figure 7 shown. It applies top-down and bottom-up bidirectional multi-scale feature fusion, enhancing the feature information fusion between the same network layers and the feature information transmission between different layers, enabling the network to have a higher important feature purification efficiency. Since different input features have different resolutions and different contributions to the output features, learnable weights are introduced in the network to learn and update the weights of different input features, so that important detailed features are better retained after fusion. The introduction of BiFPN and feature channel fusion can obtain deep semantic information and retain more location information without sacrificing computational complexity. While increasing the detection accuracy, it can still ensure the speed and real-time performance of network applications.

[0043] In the third stage, the prediction network predicts the input feature map and outputs the results. Its structure includes some convolutional layers, pooling layers, fully connected layers, etc., and outputs feature map sizes of 20×20, 40×40, and 80×80 respectively, corresponding to targets of 32×32, 16×16, and 8×8 pixels.

[0044] (3) For S3: Train the tram entering ladder detection model.

[0045] Input the processed data set into the constructed model for training. Select the SGDM optimizer to implement the iterative update of network parameters. Set the batch_size of network training to 32, the initial learning rate to 0.001, the momentum to 0.937, the epoch to 100, and the input size to 640×640.

[0046] (4) For S4: Input the image to be detected into the trained tram entering ladder detection model for testing.

[0047] Input the image to be detected into the trained model. After operations such as feature extraction, feature fusion, and prediction in the network, the final detection result is obtained.

[0048] The above specific implementation manners are only used to illustrate the technical solutions of the present invention, rather than limiting it. Those skilled in the art should understand that the above embodiments do not limit the present invention in any form. Any similar technical solutions obtained by using equivalent replacements or equivalent transformations and the like all fall within the protection scope of the present invention.

Claims

1. A deformable target detection method based on an adaptive feature extraction network and an attention mechanism, characterized in that: The following steps are involved: Step 1: Build a rich tram entry dataset; Construct a tram entry dataset. The image data of the tram entry dataset comes from public datasets, web crawlers, a community elevator surveillance video, and manual photography. The dataset contains a total of 6673 images. In addition to images of different angles, it also includes some images that are exposed due to lighting, and images that are blocked by people and objects. Step 2: Build a tram entry detection model; Design a detection model suitable for deformable targets. The detection model is designed based on the YOLOv5 model and consists of three parts: backbone network, neck network and prediction network. Phase 1: The main task of the backbone network in the detection model is to extract the features of the tram entering the elevator image input in step 1, and pass the features to the subsequent network layers for tram entering the elevator detection; the backbone network is composed of Focus structure, C3 structure and SPP structure. Focus fuses the image information to the channel through slicing operation, and then extracts the tram target features, so that the feature extraction is more sufficient; The C3 structure divides the input feature map into two sub-maps, one of which is sent to a convolution layer consisting of three consecutive 3×3 convolution operations for feature extraction, and the other sub-map is directly concat-operated with the result of the convolution; the SPP structure uses window pooling of different sizes to improve the extraction effect of tram features; The residual structure in the backbone network C3 structure is fused with deformable convolution and changed into CD2Bottleneck module, which is integrated into the backbone network to replace the original C3 structure to adapt to the changes in the detailed characteristic information caused by the appearance and morphological changes of different targets. The weighted attention mechanism SimAM is added to improve the ability to extract the detailed features of the target, enhance the feature learning ability of the deep network, obtain the tram feature map with stronger semantic information, and enable the tram entering the elevator detection model to accurately extract the key feature information of the detection target and accurately perform the detection task of deformable targets. The second stage: the neck network adopts the BiFPN structure improved on the PANet structure, and fuses the different feature layers of the tram extracted by the backbone network; The PANet structure consists of two feature pyramids. One is a top-down feature pyramid structure. Through a series of convolution and downsampling operations, multi-scale tram feature maps are extracted from the detection head. The feature maps of the top-down path have higher resolution and semantic information. The other is a bottom-up inverted pyramid structure. Through a series of convolution and upsampling operations, multi-scale tram feature maps are extracted from the backbone network. The tram feature maps have different resolutions and semantic information. Then the bottom-up and top-down feature maps are fused layer by layer through the concat operation, so that features of different resolutions can complement and fuse each other. The BiFPN structure applies top-down and bottom-up bidirectional multi-scale feature fusion to enhance the tram feature information fusion between the same network layers and the tram feature information transmission between different layers. The third stage: The prediction network consists of a convolutional layer, a fully connected layer, and an activation function, which extracts features and compresses the input tram feature map. The prediction network has three decoupling heads, which are used to predict target boxes of different scales. Each decoupling head contains a series of convolutional layers and fully connected layers, which are used to extract features from the input tram feature map and predict the target category and bounding box. Step 3: Train the tram entry detection model; Step 4: Input the image to be recognized into the trained tram-entering-elevator model for detection and verify the detection effect.

2. The method for deformable target detection based on adaptive feature extraction network and attention mechanism according to claim 1, characterized in that: The backbone network extracts features from the input image to be detected. The specific steps are as follows: The Focus slicing operation splits the input feature map into four smaller feature maps, performs a convolution operation on each small feature map, and finally stacks these small feature maps together to form an output feature map; the input image of size 640×640×3 is sampled every other pixel to obtain four independent feature layers, each of which is 320×320×3 in size; the four independent feature layers are stacked together to form a feature map of size 320×320×12, in which the width and height information is integrated into the channel information, and the number of input channels is quadrupled.

3. The method for deformable target detection based on adaptive feature extraction network and attention mechanism according to claim 2, characterized in that: The Conv operation in the backbone network consists of three parts: convolution, Batch Normalization, and SiLu activation layer.

4. The method for deformable target detection based on adaptive feature extraction network and attention mechanism according to claim 3, characterized in that: The C2D operation consists of two parts. The first part uses the above-specified multiple Bottleneck stacks, and the other part only passes through a basic convolution module, and finally the two branches are concat operated. DCNv2 is introduced in the Bottleneck module, and DCNv2 introduces a deformable convolution layer. The construction of DCNv2 includes two key steps. The first is the offset generation stage, in which the convolution operation generates the sampling point offset of the convolution kernel on the input feature map, covering the x and y directions. The second is the sampling and convolution stage, which uses the input feature map and the calculated offset for bilinear interpolation to accurately determine the sampling point position of the convolution kernel on the input feature map. Through these sampling points, the convolution operation is finally performed to realize the processing of the feature map.

5. The method for deformable target detection based on adaptive feature extraction network and attention mechanism according to claim 4, characterized in that: The residual structure Bottleneck in the original module of the backbone network is fused with deformable convolution and changed into a CD2Bottleneck module; The SimAM attention mechanism added to the structure is a 3D attention mechanism that calculates the three-dimensional attention weights and can make full use of the information of three-dimensional data to capture the detailed features of the target; The SPP operation first uses a standard convolution module to halve the input channels, and then uses three maximum pooling operations with a pooling kernel of 5 to connect them in a serial and parallel manner. The results of the three maximum pooling operations are concat with the data without pooling operations. The number of channels after the final merger is twice the original number. The operation is completed after convolution, batch normalization, and ReLu activation function operations.

6. The method for deformable object detection based on adaptive feature extraction network and attention mechanism according to claim 1, characterized in that: The neck network extracts features from the input image to be detected and fuses information from different feature map levels to improve the target detection performance; the BiFPN network structure adopted by the neck network applies top-down and bottom-up bidirectional multi-scale feature fusion to enhance the feature information fusion between the same network layers and the feature information transmission between different layers.

7. The method for deformable object detection based on adaptive feature extraction network and attention mechanism according to claim 1, characterized in that: The prediction network predicts the input feature map and outputs the result, including convolutional layer, pooling layer and fully connected layer, which output feature map sizes of 20×20, 40×40 and 80×80 respectively, corresponding to targets of 32×32, 16×16 and 8×8 pixels.

8. The method for deformable object detection based on adaptive feature extraction network and attention mechanism according to claim 1, characterized in that: When training the network model, the SGDM optimizer is selected to implement iterative updates of network parameters. The batch_size of the network training is set to 32, the initial learning rate is set to 0.001, the momentum is 0.937, the epoch is set to 100, and the input size is 640×640. When training the tram-entering-elevator detection model, the adaptive learning rate and early stopping mechanism are introduced to find a balance between training speed and accuracy. Adaptive learning rate dynamically adjusts the learning rate based on changes in loss to speed up convergence and improve model accuracy.

9. The method for deformable object detection based on adaptive feature extraction network and attention mechanism according to claim 1, characterized in that: The image to be detected is input into the trained model, and after feature extraction, feature fusion and prediction operations are performed in the network, the final detection result is obtained.

Images