Three-dimensional target detection method based on monocular vision

Abstract

The invention provides a three-dimensional target detection method based on monocular vision, and belongs to the field of three-dimensional target detection for automatic driving, and the method comprises the following steps: designing an improved backbone network Faster Net +, carrying out image feature extraction, constructing a multi-dimensional feature adaptive fusion module, and adaptively selecting and fusing high-dimensional and low-dimensional features; a feature enhancement attention module is introduced on a multi-scale feature layer extracted by the feature pyramid network, interaction between feature channels and correlation between different coordinates are considered at the same time, a target area is highlighted, and irrelevant background information is inhibited; introducing an effective target detection head network according to the three-dimensional target detection network model, and performing parameter learning on the network model by using a training data set; after training is finished, a test image is input, and the positions and categories of different types of targets in the image are determined by using the three-dimensional target detection network model.

Description

Technical Field

[0001] The present invention relates to technical fields such as computer vision and deep learning, and particularly to a three-dimensional target detection method based on monocular vision. Background Art

[0002] With technological advancements and growing demand for intelligence, technologies such as autonomous driving, intelligent transportation, and robotics have become key areas of development. Object detection, as a key technology, has become increasingly important. In particular, three-dimensional object detection, which provides precise information such as object type, size, and pose, has become a hot topic of research. Commonly used sensors for three-dimensional object detection include lidar, monocular cameras, and binocular cameras. While lidar offers high accuracy, its high cost and sensitivity to complex weather conditions limit its widespread application. Consequently, vision-based three-dimensional object detection methods, particularly monocular vision technology, have garnered increasing attention. These methods require only a single camera and its calibration for three-dimensional detection. While technically challenging, their low cost and ease of implementation hold significant potential in applications such as autonomous driving. Therefore, in-depth research and continuous improvement of the accuracy of monocular three-dimensional object detection have significant academic value and practical application prospects.

[0003] There are three main approaches to monocular 3D object detection: depth map methods, lidar-assisted methods, and direct regression methods. Depth map methods rely on pre-acquired pixel-level depth maps, and their prediction errors can further impact the performance of monocular 3D object detection models. LiDAR-assisted methods use lidar point cloud information to aid model training, but due to the time-consuming training process, they often struggle to meet real-time requirements. Alternatively, there are methods based on direct regression that improve algorithm performance by leveraging geometric prior knowledge and uncertainty modeling of depth estimation. Overall, current monocular 3D object detection methods primarily focus on optimizing detector frameworks and improving detection accuracy through various regression strategies. However, these methods often overlook the problems of false and missed detections caused by object scale differences and occlusion in complex road scenes, thus limiting improvements in detection performance. Summary of the Invention

[0004] The present invention provides a three-dimensional target detection method based on monocular vision to overcome the above technical problems. To achieve the above object, the technical solution adopted by the present invention is:

[0005] A three-dimensional target detection method based on monocular vision includes the following steps:

[0006] Acquire the image to be detected;

[0007] Build a 3D object detection network model;

[0008] The 3D object detection network model is trained based on the training set data, and parameters are learned to obtain a trained 3D object detection network model, which is used to detect the position and category of different types of objects in the image;

[0009] The image to be detected is input into the trained 3D object detection network model to detect the position and category of different types of objects in the image.

[0010] Furthermore, the three-dimensional object detection network model includes:

[0011] Improved backbone network FasterNet+: used to extract local features and detail information from images;

[0012] Multi-dimensional feature adaptive fusion module: used to adaptively select and fuse high-dimensional features and low-dimensional features based on the features extracted by the improved backbone network FasterNet+;

[0013] Feature Pyramid Network: used for multi-scale feature extraction of the fused high-dimensional features and low-dimensional features output by the multi-dimensional feature adaptive fusion module;

[0014] Feature Enhancement Attention Module: This module is used to highlight the target area, suppress irrelevant background information, and improve the perception of features at each scale based on the features extracted by the feature pyramid network while considering the interaction between feature channels and the correlation between different coordinates.

[0015] Detection head network: It is used to enhance the information output by the attention module based on the features, extract information from feature maps of different scales, and obtain the final detection results after different regression branches.

[0016] Furthermore, the improved backbone network FasterNet+ includes four hierarchical levels, each of which is preceded by an embedding layer or a merging layer for performing spatial downsampling and expanding the number of channels, and each hierarchical level contains multiple FasterNet+ blocks;

[0017] A combination of multiple convolutional layers is used as the embedding layer to retain more local information and improve the fine-grained feature extraction capability without sacrificing accuracy. Specifically, a 3×3 regular convolutional layer with a stride of 2 is used for preliminary processing, and then a 3×3 deep convolutional layer is used to extract rich low-level features channel by channel. A 1×1 convolutional layer is used for feature fusion, and another 3×3 deep convolutional layer is introduced to achieve downsampling. At the same time, feature transfer is enhanced through skip connections.

[0018] The merging layer uses a regular 2×2 convolution with a stride of 2;

[0019] Each FasterNet+ block consists of two point-by-point convolutional layers plus a 3×3 partial convolutional layer in the middle. A normalization layer and an activation layer are added after each intermediate layer to maintain feature diversity and effectively reduce latency.

[0020] Furthermore, the multi-dimensional feature adaptive fusion module adaptively selects and fuses high-dimensional features and low-dimensional features for the features extracted by the improved backbone network FasterNet+. The implementation process is as follows:

[0021] S41: Processing low-dimensional features through group convolution;

[0022] Process high-dimensional features through bilinear interpolation to achieve preliminary alignment with the current layer features;

[0023] S42: Divide the three features of high-order features, low-dimensional features and current layer features into four equal parts in the channel dimension, and obtain the feature l of the i-th division of low-dimensional, high-dimensional and current layer features. i 、h i and u i ;

[0024] To u i Use the sigmoid function to calculate the fusion coefficients k and 1-k respectively, and calculate l according to the fusion coefficients. i and h i Integrate and then splice in the channel dimension to obtain selective fusion features;

[0025] S43: The selective fusion features and the original input features are added after 1×1 convolution, and then processed by batch normalization and activation function to finally output a new feature map.

[0026] Furthermore, the feature-enhanced attention module includes:

[0027] The first branch: contains a 1×1 convolution and a global coordinate attention module;

[0028] The second branch: contains only one 1×1 convolution,

[0029] The outputs of these two branches are combined and then go through a 1×1 convolution;

[0030] The implementation process of the global coordinate attention module is as follows:

[0031] S51: For a given input feature map of size [C, H, W], three pooling operations are first performed: adaptive average pooling with pooling kernels of (H, 1) and (1, W) is used to encode the horizontal and vertical coordinates, respectively, to obtain feature maps of size [C, H, 1] and [C, 1, W]. At the same time, global average pooling is used to globally encode the channel spatial information, resulting in a feature map of size [C, 1, 1].

[0032] S52: The three pooled feature maps are concatenated and processed through 1×1 convolution, batch normalization and activation function;

[0033] S53: Split and transform the processed feature map and then perform 1×1 convolution and sigmoid operations respectively;

[0034] The three obtained feature maps are then multiplied with the original input feature map to enhance its representation ability.

[0035] Furthermore, the target detection head network uses the head network of FCOS3D, which includes two parts: classification branch and regression branch. Each branch consists of four convolutional layers with shared parameters. These convolutional layers are used to extract information from feature maps of different scales.

[0036] A three-dimensional target detection device based on monocular vision, comprising:

[0037] Acquisition module: used to acquire the image to be detected;

[0038] Building module: used to build a 3D target detection network model;

[0039] Training module: used to train the 3D object detection network model based on the training set data, perform parameter learning, and obtain a trained 3D object detection network model for detecting the position and category of different types of objects in the image;

[0040] Implementation module: used to input the image to be detected into the trained 3D object detection network model to detect the position and category of different types of objects in the image.

[0041] A computer device comprises: a processor and a memory, wherein the memory stores a program module, and wherein the program module runs on the processor to implement any one of the methods described above.

[0042] A readable storage medium stores a program module, wherein the program module can implement any of the methods described above when executed in a processor.

[0043] The beneficial effects of the present invention are:

[0044] The present invention provides a three-dimensional target detection method based on monocular vision, which realizes a fast and high-precision target detection method under the nuScenes dataset;

[0045] This paper uses FasterNet+ as the backbone network and optimizes the embedding layer and module structure to better capture detailed information and improve the network's overall performance. Furthermore, a multi-dimensional feature adaptive fusion module is designed. By adaptively selecting and fusing high- and low-dimensional features, it effectively addresses the problem of high-dimensional features losing information about small targets and low-dimensional features lacking contextual information. Furthermore, a feature-enhanced attention module is introduced to highlight target areas, further improving the network's accuracy in target localization and classification.

[0046] By using the official evaluation metrics of the nuScenes dataset to analyze the detection effect on the validation set and comparing it with the detection performance of the classic 3D object detection model on the same dataset, it was found that the detection accuracy of the proposed method has been significantly improved. It can effectively detect objects of different types and scales and exhibits stronger robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 any creative work.

[0048] Figure 1 It is the overall flow chart of the present invention;

[0049] Figure 2 It is the network framework of the present invention;

[0050] Figure 3 Schematic diagram of the improved backbone network FasterNet+ structure of the present invention;

[0051] Figure 4 It is a schematic diagram of the structure of the multi-dimensional feature adaptive fusion module of the present invention;

[0052] Figure 5 Schematic diagram of the feature-enhanced attention module structure of the present invention;

[0053] Figure 6 Schematic diagram of the 3D target detection results of the nuScenes dataset using the present invention, where (a) compares the detection results of different methods during the day, (b) compares the detection results of different methods on rainy days, and (c) compares the detection results of different methods at night. DETAILED DESCRIPTION

[0054] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0055] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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. The following description of at least one exemplary embodiment is actually only illustrative and is in no way intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0056] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, singular form schematic diagrams include plural forms. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0057] Unless otherwise specifically stated, the relative arrangement of the parts and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be clear that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The technology, methods and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but in appropriate cases, the technology, methods and equipment should be considered as part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary and not as a limitation. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that similar numbers and letters represent similar items in the following figures, and therefore, once an item is defined in one figure, it does not need to be further discussed in subsequent figures.

[0058] In the description of the present invention, it should be understood that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "horizontal, vertical, vertical, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific direction or be constructed and operated in a specific direction. Therefore, they cannot be understood as limiting the scope of protection of the present invention: the directional words "inside and outside" refer to the inside and outside relative to the outline of each component itself.

[0059] In addition, it should be noted that the use of terms such as "first" and "second" to limit components is only for the convenience of distinguishing the corresponding components. Unless otherwise stated, the above terms have no special meaning and therefore cannot be understood as limiting the scope of protection of the present invention.

[0060] Figure 1 Here is the flow chart of this method:

[0061] A three-dimensional target detection method based on monocular vision includes the following steps:

[0062] S1: Acquire the image to be detected;

[0063] S2: Build a 3D object detection network model;

[0064] S3: Training the 3D object detection network model based on the training set data, performing parameter learning, and obtaining a trained 3D object detection network model for detecting the position and category of different types of objects in the image;

[0065] S4: Input the image to be detected into the trained 3D object detection network model to detect the position and category of different types of objects in the image.

[0066] Steps S1, S2, S3 and S4 are performed sequentially.

[0067] Figure 2 This is the network architecture of this method:

[0068] The three-dimensional target detection network model includes:

[0069] Improved backbone network FasterNet+: used to extract local features and detail information from images;

[0070] Multi-dimensional feature adaptive fusion module: used to adaptively select and fuse high-dimensional features and low-dimensional features based on the features extracted by the improved backbone network FasterNet+;

[0071] Feature Pyramid Network: used for multi-scale feature extraction of the fused high-dimensional features and low-dimensional features output by the multi-dimensional feature adaptive fusion module;

[0072] Feature Enhancement Attention Module: This module is used to highlight the target area, suppress irrelevant background information, and improve the perception of features at each scale based on the features extracted by the feature pyramid network while considering the interaction between feature channels and the correlation between different coordinates.

[0073] Detection head network: It is used to enhance the information output by the attention module based on the features, extract information from feature maps of different scales, and obtain the final detection results after different regression branches.

[0074] The improved backbone network, FasterNet+, builds on FasterNet by optimizing the embedding layer and block structure. By introducing a combination of multiple convolutional layers within the embedding layer to capture more detailed information, and adjusting the order of convolutional layers within the FasterNet block, the overall performance of the network is effectively improved.

[0075] Furthermore, an improved backbone network FasterNet+ is designed according to step S1 to perform image feature extraction;

[0076] Specifically, if Figure 3 As shown, the improved backbone network FasterNet includes four hierarchical levels, each of which is preceded by an embedding layer or a merging layer for performing spatial downsampling and expanding the number of channels, and each hierarchical level contains multiple FasterNet+ blocks;

[0077] A combination of multiple convolutional layers is used as the embedding layer to retain more local information and improve the fine-grained feature extraction capability without sacrificing accuracy. Specifically, a 3×3 regular convolutional layer with a stride of 2 is used for initial processing. A 3×3 deep convolutional layer (DWConv) is then used to extract rich low-level features channel by channel. A 1×1 convolutional layer is used for feature fusion, and another 3×3 deep convolutional layer (DWConv) is introduced to achieve downsampling. Skip connections are also used to enhance feature transfer. This multi-layer convolutional structure provides a more flexible downsampling method, allowing the size of the feature map to be gradually controlled, thereby improving the network's adaptability to complex scenes and small objects.

[0078] The merging layer uses a regular 2×2 convolution with a stride of 2;

[0079] Each FasterNet+ block consists of two point-by-point convolutional layers (PWConv, or 1×1 convolutional layers) plus a central 3×3 partial convolutional layer (PConv). Normalization and activation layers are added after each intermediate layer to maintain feature diversity and effectively reduce latency.

[0080] Furthermore, the multi-dimensional feature adaptive fusion module adaptively selects and fuses high-dimensional features and low-dimensional features for the features extracted by the improved backbone network FasterNet+. The implementation process is as follows:

[0081] Specifically, if Figure 4 As shown, the implementation process of the multi-dimensional feature adaptive fusion module is as follows:

[0082] S41: Processing low-dimensional features through group convolution;

[0083] Process high-dimensional features through bilinear interpolation to achieve preliminary alignment with the current layer features;

[0084] S42: Divide these three features into four equal parts in the channel dimension to obtain the i-th divided features l of low-dimensional, high-dimensional and current layer features. i 、h i and u i . i Use the sigmoid function to calculate the fusion coefficients k and 1-k respectively, integrate li and hi according to the fusion coefficients, and then splice them in the channel dimension to obtain the selective fusion features;

[0085] S43: In order to enhance the feature extraction capability of the network, the selective fusion features and the original input features are added after 1×1 convolution, and then processed by batch normalization and activation function to finally output a new feature map.

[0086] Furthermore, according to step S2, a feature enhancement attention module is introduced on the multi-scale feature layer extracted by the feature pyramid network, which considers the interaction between feature channels and the correlation between different coordinates, highlights the target area, and suppresses irrelevant background information;

[0087] Specifically, if Figure 5 As shown, the feature enhanced attention module

[0088] The feature-enhanced attention module includes:

[0089] The first branch: contains a 1×1 convolution and a global coordinate attention module;

[0090] The second branch: contains only one 1×1 convolution,

[0091] The outputs of these two branches are combined and then go through a 1×1 convolution;

[0092] The implementation process of the global coordinate attention module is as follows:

[0093] S51: For a given input feature map of size [C, H, W], three pooling operations are first performed: adaptive average pooling with pooling kernels of (H, 1) and (1, W) is used to encode the horizontal and vertical coordinates, respectively, to obtain feature maps of size [C, H, 1] and [C, 1, W]. At the same time, global average pooling is used to globally encode the channel spatial information, resulting in a feature map of size [C, 1, 1].

[0094] S52: The three pooled feature maps are concatenated and processed through 1×1 convolution, batch normalization and activation function;

[0095] S53: The processed feature maps are split and dimensionally transformed, and then subjected to 1×1 convolution and sigmoid operations respectively. The three obtained feature maps are then multiplied with the original input feature map to enhance its representation capability.

[0096] The target detection head network uses the head network of FCOS3D, which includes two parts: classification branch and regression branch. Each branch consists of four convolutional layers with shared parameters. These convolutional layers are used to extract information from feature maps of different scales.

[0097] Furthermore, the experimental dataset used, nuScenes, is a multimodal dataset widely used in autonomous driving research, released by Boston-based Motional. This dataset aims to provide comprehensive data support for the perception, decision-making, and planning algorithms of autonomous driving systems, with a particular focus on application scenarios in complex urban environments. The nuScenes dataset contains 1,000 rich and diverse scenes, each lasting approximately 20 seconds, covering different types of traffic scenarios and road conditions, such as city streets, highways, and parking lots. The uniqueness of this dataset lies in its fusion of multiple sensors, providing rich data from multiple sensors, including approximately 1.4 million camera images, 390,000 LiDAR scans, 1.4 million radar scans, and 1.4 million object bounding boxes in 40,000 keyframes. The table details the sensor parameters used in the dataset. The nuScenes dataset contains 10 common object categories: car, truck, trailer, bus, construction vehicle, bicycle, motorcycle, pedestrian, traffic cone, and barrier. These categories comprehensively cover dynamic traffic participants and static obstacles in urban road environments, providing essential training and validation data for core autonomous driving tasks such as environmental perception, path planning, and behavior prediction.

[0098] Specifically, if Figure 6 As shown in the figure, (a) is the comparison of detection results of different methods during the day, (b) is the comparison of detection results of different methods on rainy days, and (c) is the comparison of detection results of different methods at night.

[0099] The trained model can accurately extract the feature description of the target in the image and automatically determine the location and category information of the target.

[0100] Furthermore, in order to objectively compare the detection effects of different network models, the official evaluation indicators of the nuScenes dataset were used to analyze the target detection effects on the validation set. The mean average precision (mAP) is calculated by averaging the average precision (AP) of each category, aiming to evaluate the overall detection ability of the network model on all categories of targets. The TP indicators include: average translation error (ATE), average scale error (ASE), average orientation error (AOE), average velocity error (AVE), and average attribute error (AAE). Based on these indicators, the average TP indicator (mTP) of all categories is calculated. NDS is a comprehensive indicator that comprehensively evaluates the position, size, direction, velocity and attributes of the target.

[0101] Example

[0102] This paper provides a monocular vision-based 3D object detection method that achieves rapid and high-precision object detection using autonomous driving datasets. To verify the effectiveness of the proposed monocular vision-based 3D object detection method, a comparative experiment was conducted on the nuScenes dataset with several other common 3D object detection methods.

[0103] Table 1 shows the comparison of detection accuracy of different methods on the nuScenes dataset, where the bold data shows the best detection results and the underlined data shows the suboptimal results;

[0104]

[0105] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional target detection method based on monocular vision, characterized in that: The steps include: Acquire the image to be detected; Build a 3D object detection network model; The 3D object detection network model is trained based on the training set data, and parameters are learned to obtain a trained 3D object detection network model, which is used to detect the position and category of different types of objects in the image; The image to be detected is input into the trained 3D object detection network model to detect the position and category of different types of objects in the image.

2. The three-dimensional target detection method based on monocular vision according to claim 1, characterized in that: The three-dimensional target detection network model includes: Improved backbone network FasterNet+: used to extract local features and detail information from images; Multi-dimensional feature adaptive fusion module: used to adaptively select and fuse high-dimensional features and low-dimensional features based on the features extracted by the improved backbone network FasterNet+; Feature Pyramid Network: used for multi-scale feature extraction of the fused high-dimensional features and low-dimensional features output by the multi-dimensional feature adaptive fusion module; Feature Enhancement Attention Module: This module is used to highlight the target area, suppress irrelevant background information, and improve the perception of features at each scale based on the features extracted by the feature pyramid network while considering the interaction between feature channels and the correlation between different coordinates. Detection head network: It is used to enhance the information output by the attention module based on the features, extract information from feature maps of different scales, and obtain the final detection results after different regression branches.

3. The method for three-dimensional target detection based on monocular vision according to claim 1, characterized in that: The improved backbone network FasterNet+ consists of four hierarchical levels, each of which is preceded by an embedding layer or a merging layer for performing spatial downsampling and expanding the number of channels. Each hierarchical level contains multiple FasterNet+ blocks. A combination of multiple convolutional layers is used as the embedding layer to retain more local information and improve the fine-grained feature extraction capability without sacrificing accuracy. Specifically, a 3×3 regular convolutional layer with a stride of 2 is used for preliminary processing, and then a 3×3 deep convolutional layer is used to extract rich low-level features channel by channel. A 1×1 convolutional layer is used for feature fusion, and another 3×3 deep convolutional layer is introduced to achieve downsampling. At the same time, feature transfer is enhanced through skip connections. The merging layer uses a regular 2×2 convolution with a stride of 2; Each FasterNet+ block consists of two point-by-point convolutional layers plus a 3×3 partial convolutional layer in the middle. A normalization layer and an activation layer are added after each intermediate layer to maintain feature diversity and effectively reduce latency.

4. The method for three-dimensional target detection based on monocular vision according to claim 1, wherein: The multi-dimensional feature adaptive fusion module adaptively selects and fuses high-dimensional features and low-dimensional features based on the features extracted by the improved backbone network FasterNet+. The implementation process is as follows: S41: Processing low-dimensional features through group convolution; Process high-dimensional features through bilinear interpolation to achieve preliminary alignment with the current layer features; S42: Divide the three features of high-order features, low-dimensional features and current layer features into four equal parts in the channel dimension, and obtain the i-th divided features l of low-dimensional, high-dimensional and current layer features. i 、h i and u i ; To u i Use the sigmoid function to calculate the fusion coefficients k and 1-k respectively, and calculate l according to the fusion coefficients. i and h i Integrate and then splice in the channel dimension to obtain selective fusion features; S43: The selective fusion features and the original input features are added after 1×1 convolution, and then processed by batch normalization and activation function to finally output a new feature map.

5. The method for three-dimensional target detection based on monocular vision according to claim 1, characterized in that: The feature-enhanced attention module includes: The first branch: contains a 1×1 convolution and a global coordinate attention module; The second branch: contains only one 1×1 convolution, The outputs of these two branches are combined and then go through a 1×1 convolution; The implementation process of the global coordinate attention module is as follows: S51: For a given input feature map of size [C, H, W], three pooling operations are first performed: adaptive average pooling with pooling kernels of (H, 1) and (1, W) is used to encode the horizontal and vertical coordinates, respectively, to obtain feature maps of size [C, H, 1] and [C, 1, W]. At the same time, global average pooling is used to globally encode the channel spatial information, resulting in a feature map of size [C, 1, 1]. S52: The three pooled feature maps are concatenated and processed through 1×1 convolution, batch normalization and activation function; S53: Split and transform the processed feature map and then perform 1×1 convolution and sigmoid operations respectively; The three obtained feature maps are then multiplied with the original input feature map to enhance its representation ability.

6. The method for three-dimensional target detection based on monocular vision according to claim 1, characterized in that: The target detection head network uses the head network of FCOS3D, which includes two parts: classification branch and regression branch. Each branch consists of four convolutional layers with shared parameters. These convolutional layers are used to extract information from feature maps of different scales.

7. A three-dimensional target detection device based on monocular vision, characterized in that: include: Acquisition module: used to acquire the image to be detected; Building module: used to build a 3D target detection network model; Training module: used to train the 3D object detection network model based on the training set data, perform parameter learning, and obtain a trained 3D object detection network model for detecting the position and category of different types of objects in the image; Implementation module: used to input the image to be detected into the trained 3D object detection network model to detect the position and category of different types of objects in the image.

8. A computer device comprising: A processor and a memory, wherein the memory stores a program module, wherein the program module runs on the processor to implement the method according to any one of claims 1 to 6.

9. A readable storage medium storing a program module, characterized in that: The program module is run in a processor to implement the method according to any one of claims 1 to 6.

Images