Unmanned aerial vehicle target detection method based on feature fusion and detection head collaborative optimization

Abstract

The invention discloses an unmanned aerial vehicle target detection method based on feature fusion and detection head collaborative optimization, and the method comprises the steps: obtaining a multi-scale feature map through employing a backbone network formed by stacking a convolution module CBS and a residual module C2f, sequentially transmitting the feature maps of various scales to a multi-level feature fusion module (ML-FFM), carrying out the feature fusion from a deep layer to a shallow layer, and carrying out the detection of a target target. Semantic information of shallow layer features is enhanced; inputting the fused feature map into a refined feature enhancement module (RFEM), eliminating position conflicts and redundant information based on a space-time attention mechanism, and optimizing details and positioning capability of deep features; and the enhanced feature map is sent to a prediction head based on Transform for detection, and target classification and positioning of the aerial image of the unmanned aerial vehicle are completed. According to the method, the problems that the unmanned aerial vehicle aerial image is large in scale change, and small targets are high in proportion and dense in distribution are solved, so that the precision of target detection in the unmanned aerial vehicle aerial image is improved. The modules proposed in the method can significantly reduce the calculation amount, and are suitable for the real-time detection scene of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to a drone target detection method based on feature fusion and detection head collaborative optimization, specifically a target detection method for a drone with a high proportion of small targets and dense distribution under the perspective of the drone, belonging to the technical field of target detection. Background Art

[0002] As an important research direction of computer vision and digital image processing, target detection has been widely used in intelligent assisted driving, industrial inspection and other fields. Through target detection technology, computer vision algorithms can significantly reduce manpower consumption and provide more accurate services. Target detection uses the basic target classification network to further realize the automatic recognition of objects in pictures or videos, and has become a research hotspot in the field of computer vision and artificial intelligence. In drone applications, target detection gives drones the ability to autonomously perceive, analyze and make decisions, making them show a wide range of application prospects in traffic monitoring, power inspection, water inspection and disaster relief. Drones have a unique aerial perspective and high maneuverability, which can achieve efficient data collection, especially in traffic monitoring, showing advantages such as fast response and wide field of view. However, most of the current target detection research focuses on ground video monitoring, and target detection in drone-viewed images has not been fully explored. The accuracy of existing algorithms on drone images is generally low.

[0003] Object detection from the perspective of drones faces many challenges. First, the flexible flight altitude and variable perspective of drones cause significant changes in the distance of the objects being photographed, and the scale of the objects in the image is extremely uneven. In addition, the wide field of view of drones covers a large number of targets and also includes a large amount of background, resulting in a clustered distribution of targets in the image: some areas are densely populated with occlusion problems, while some areas are sparse or even without targets, increasing the waste of computing resources. Compared with ground-based images, there are more foreground targets in the perspective of drones, but they account for a small proportion of the image, resulting in an imbalance of positive and negative samples, and the model is easily disturbed by background noise, making it difficult to accurately focus on the target. Furthermore, since the image resolution of drone data sets is usually high, processing high-resolution images requires a lot of video memory and computing resources, while the chips of drone platforms are limited by low power consumption and cannot afford high computing power. In addition, real-time processing of high-resolution images requires low latency, so a balance must be struck between computing power and detection accuracy. Summary of the invention

[0004] Purpose of the invention: In view of the problems and shortcomings of the prior art, it is necessary to fully extract the feature information of drone aerial images, enhance the features of small targets by feature fusion, and thus improve the accuracy of drone vehicle target detection. However, small targets account for a large proportion in drone aerial images, and targets are densely distributed in a small area. If the entire high-resolution image is processed, a large amount of computing resources will be wasted. In order to overcome this problem, the present invention provides a drone perspective target detection method based on feature fusion and detection head collaborative optimization. By extracting multi-scale features of drone aerial images and effectively fusing contextual information of features at different levels, a sparse detection head is introduced to accurately locate small targets, which not only greatly improves the accuracy and efficiency of drone perspective target detection, but also reduces the computational complexity of the model, which is of great significance for real-time drone target detection.

[0005] Technical solution: A UAV target detection method based on feature fusion and detection head collaborative optimization, including the following steps:

[0006] (1) Feature extraction is performed on the input drone aerial images, and a multi-scale feature map is obtained using a backbone network stacked by a convolutional module CBS and a residual module C2f;

[0007] (2) The extracted multi-scale features are sent to the multi-level feature fusion module ML-FFM to fuse the high, medium and low-level features from deep to shallow layers to obtain the fused feature map, thereby strengthening the semantic features of the shallow feature map;

[0008] (3) The fused feature map is input into the refined feature enhancement module (RFEM), which eliminates position conflicts and redundant information based on the spatiotemporal attention mechanism, optimizes the details and positioning ability of the deep features, and obtains the enhanced feature map;

[0009] (4) The enhanced feature map is sent to the Transformer-based prediction head for detection to complete the target classification and positioning of the UAV aerial image.

[0010] The multi-level feature fusion module ML-FFM is used to fuse high, medium and low-level features from deep to shallow layers, and the semantic features of shallow feature maps are strengthened. The specific steps are as follows:

[0011] (2.1) First, convolve the high-level feature map to extract high-level semantic features, normalize the convolved features to stabilize the training process, and then apply the SiLU activation function to obtain the processed high-level feature map F′ high :

[0012]

[0013] where F′ high Represents the processed high-level feature map, Fhigh represents the original high-level feature map, represents the upsampling operation, represents the SiLU activation function, represents batch normalization, Represents a convolution operation.

[0014] (2.2) Then, a depth-separable convolution is performed on the low-level feature map, which effectively reduces the amount of computation while retaining important spatial information, and the processed low-level feature map F′ is obtained. low :

[0015]

[0016] where F′ low Represents the processed low-level feature map, F low represents the original low-level feature map, represents the depthwise convolution, Represents point-wise convolution.

[0017] (2.3) Finally, the upsampled high-level feature map, the original mid-level feature map, and the low-level feature map after depth-separable convolution are concatenated, and the fused features are extracted through the residual module C2f to obtain F fusion :

[0018]

[0019] Among them, F fusion Represents the output of the multi-level feature fusion module ML-FFM, represents the residual module C2f, Indicates splicing by channel, F′ high Represents the high-level feature map after convolution module processing, F mid Represents the original mid-level feature map, F′ low Represents the low-level feature map after depth-wise separable convolution.

[0020] (2.4) Repeat steps (2.1) to (2.3), process the feature maps of different levels from top to bottom, complete the fusion of multi-scale features, and strengthen the semantic information of shallow feature maps, and finally obtain the extremely small scale feature f′ after feature fusion tiny , low-level features f′ low , intermediate feature f′ mid and high-level features f′ high .

[0021] The refined feature enhancement module (RFEM) eliminates position conflicts and redundant information based on the spatiotemporal attention mechanism, and optimizes the details and positioning capabilities of deep features. The specific steps are as follows:

[0022] (3.1) The extremely small scale feature f′ after feature fusion tiny The image is sent to the C3STR module based on SwinTransformer, and the spatial and semantic expression capabilities of the feature map are enhanced through the self-attention mechanism. SwinTransformer can focus on specific important areas while maintaining global features, which is suitable for the refined processing of low-level feature maps. The C3TR module used to process the lowest-level features is replaced by the C3STR module based on SwinTransformer, which outputs the extremely small-scale features f″ after feature enhancement. tiny .

[0023] (3.2) The deep separable convolution of the lowest level multi-level feature fusion module ML-FFM is used to extract the class-independent mask map of the dense area of ​​small objects; this mask map can indicate the approximate distribution area of ​​small objects in the feature map. First, a deep separable convolution with a convolution kernel size of 13×13 is used to expand the receptive field and capture the distribution information of small objects in high-resolution images; then, the features are mapped to a single channel through a 1×1 convolution, and the output of this channel will be used as the basis for generating the class-independent mask map. The single-channel feature map output by the 1×1 convolution is batch normalized to stabilize the feature distribution and ensure the consistency of the numerical range in subsequent processing; finally, the Sigmoid activation function is used to limit the value of the feature map to [0, 1].

[0024] (3.3) The feature obtained in step (3.1) is enhanced to obtain the smallest scale feature f″ tiny The low-level features f′ output by the multi-level feature fusion module ML-FFM low The splicing is then performed and the splicing result is sent to the Transformer-based C3TR module to extract contextual semantic feature information.

[0025] (3.4) The output of the Transformer-based C3TR module is used as the input feature f input It is sent to the lightweight convolutional attention module CBAM, which will derive the attention map along the two independent dimensions of channel and space, and then combine the attention map with the input feature f input The CBAM module can adaptively enhance feature expression, help the model focus on useful target objects, resist conflicting and redundant information generated by feature fusion, and effectively improve multi-scale detection performance.

[0026] (3.5) The output of the CBAM module is passed through a convolution module CBS to adjust the scale of the feature map to obtain the enhanced small-scale feature f″ small , f″ smallUsed to splice the feature map output by the next multi-level feature fusion module ML-FFM module.

[0027] (3.6) The intermediate feature f′ after fusion of the features obtained in step (2.4) mid And the enhanced small-scale feature f″ obtained in step (3.5) small As the input of step (3.3), repeat steps (3.3) to (3.5) and use the refined feature enhancement module RFEM to enhance the feature maps of different levels from bottom to top to obtain the feature maps of different scales after feature enhancement, which are the extremely small scale features f″ tiny , small-scale features f″ small , mesoscale features f″ mid and large-scale features f″ big .

[0028] The Transformer-based prediction head is used to detect and complete the target classification and positioning of the drone aerial image. The specific steps are as follows:

[0029] (4.1) For the extremely small scale feature f″ in step (3.6) tiny , the target is detected through the Transformer-based sparse prediction head Sparse-TPH; Sparse-TPH only detects the target feature blocks that exceed the threshold γ in the mask image obtained in step (3.2), and for non-target feature blocks, Sparse-TPH uses sparse convolution to reduce the computational waste on the background.

[0030] (4.2) For the feature maps {f″ small , f″ mid , f″ big}, detect the target through the Transformer-based prediction head TPH of the corresponding scale.

[0031] The convolution module CBS in the UAV view target detection method consists of a 3×3 convolution layer, a batch normalization layer and a SiLU activation function.

[0032] The residual module C2f in the drone view target detection method is composed of a 1x1 convolution layer, multiple 3x3 residual blocks Bottleneck and a 1x1 convolution layer. First, the input channel is divided into two parts by 1x1 convolution, and the Bottleneck blocks are stacked layer by layer to extract deep features. Finally, the channels are fused by 1x1 convolution to achieve feature dimensionality reduction and fusion, effectively balance detail information and global semantic information, and are suitable for multi-scale feature extraction and efficient calculation.

[0033] The C3TR module in the drone-view target detection method consists of a multi-head self-attention layer and a multi-layer perceptron layer, and each sublayer has layer normalization, Dropout, and residual connection.

[0034] The multi-level feature fusion module ML-FFM in the drone-view target detection method fuses the feature maps of the low, medium and high levels from top to bottom through multi-level feature fusion. The module uses deep separable convolution to process low-level features, thereby reducing the amount of calculation while retaining spatial information. The fused feature map contains rich semantic information and balanced spatial details, further enhancing the semantic expression ability of shallow features.

[0035] The refined feature enhancement module RFEM in the UAV perspective target detection method is composed of modules such as C3STR, C3TR, and CBAM. It gradually enhances feature maps of different scales from bottom to top, eliminates conflicting and redundant information generated by feature fusion, and refines the spatial positioning and detail expression of features.

[0036] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the UAV target detection method based on feature fusion and detection head collaborative optimization as described above is implemented.

[0037] A computer-readable storage medium stores a computer program for executing the unmanned aerial vehicle target detection method based on feature fusion and detection head collaborative optimization as described above.

[0038] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0039] (1) The multi-level feature fusion module ML-FFM can fuse feature maps of low, medium and high sizes, balance the spatial information of low-level features and the semantic information of high-level features, thereby enhancing the recognition and classification capabilities of target locations and facilitating the detection and fine positioning of small targets. Among them, the depthwise separable convolution used on the underlying feature map can expand the receptive field without increasing parameters, effectively capture the precise location information of the target while maintaining a low amount of computation, and improve the efficiency of feature extraction.

[0040] (2) Refined feature enhancement module RFEM uses self-attention, contextual semantics and channel-space attention mechanisms to perform multi-level refinement of fused features through modules such as SwinTransformer, C3TR and CBAM. SwinTransformer focuses on the global semantic information of a specific area, the C3TR module extracts contextual semantic information, and the CBAM module adaptively removes redundant features, allowing the network to capture key feature areas more effectively, helping the model focus on useful target features in complex detection tasks and improving detection accuracy.

[0041] (3) In extremely small-scale object detection, the Transformer-based sparse prediction head Sparse-TPH only detects target feature areas that exceed the threshold, and uses sparse convolution in other areas. This strategy significantly reduces the computational overhead in the background area and improves the computational efficiency of the network, making it suitable for processing high-resolution drone aerial images. For small, medium, and large-scale feature maps, TPH only applies the Transformer encoder block on the low-resolution feature map at the end of the network, reducing computational and memory costs while maintaining good feature expression capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram of the network structure of a UAV target detection method based on feature fusion and detection head collaborative optimization in a specific embodiment;

[0043] Figure 2 It is a schematic diagram of the structure of the multi-level feature fusion module ML-FFM in a specific embodiment;

[0044] Figure 3 Schematic diagram of the RFEM structure of the refined feature enhancement module in a specific embodiment. DETAILED DESCRIPTION

[0045] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0046] It is known that in a certain city traffic monitoring system, drones carry high-definition cameras for high-altitude cruising and take real-time images of traffic flow on urban roads. These images may contain four types of traffic targets, namely cars, buses, trucks and pedestrians. It is necessary to extract the features of drone images through a model, distinguish the foreground and background, accurately detect and classify various types of traffic targets, distinguish different traffic flows, and achieve accurate analysis and statistics of road traffic. This method can significantly reduce the workload of manual monitoring, improve the real-time and accuracy of traffic flow estimation, and further optimize the efficiency of urban traffic management and emergency response.

[0047] Figure 1 The network structure diagram of the UAV target detection method based on feature fusion and detection head collaborative optimization is given, and the specific implementation is as follows:

[0048] (1) Feature extraction is performed on the input drone aerial images, and a multi-scale feature map is obtained using a backbone network stacked by a convolutional module CBS and a residual module C2f;

[0049] (2) The extracted multi-scale features are sent to the multi-level feature fusion module ML-FFM to fuse the high, medium and low-level features from deep to shallow layers, and strengthen the semantic features of the shallow feature map, such as Figure 2 As shown;

[0050] (2.1) First, convolve the high-level feature map to extract high-level semantic features, normalize the convolved features to stabilize the training process, and then apply the SiLU activation function to obtain the processed high-level feature map F′ high :

[0051]

[0052] where F′ high Represents the processed high-level feature map, F high represents the original high-level feature map, represents the upsampling operation, represents the SiLU activation function, represents batch normalization, Represents a convolution operation.

[0053] (2.2) Then, a depth-separable convolution is performed on the low-level feature map, which effectively reduces the amount of computation while retaining important spatial information, and the processed low-level feature map F′ is obtained. low :

[0054]

[0055] where F′ low Represents the processed low-level feature map, F lowrepresents the original low-level feature map, represents the depthwise convolution, Represents point-wise convolution.

[0056] (2.3) Finally, the upsampled high-level feature map, the original mid-level feature map, and the low-level feature map after depth-separable convolution are concatenated, and the fused features are extracted through the residual module C2f to obtain F fusion :

[0057]

[0058] Among them, F fusion Represents the output of the multi-level feature fusion module ML-FFM, represents the residual module C2f, Indicates splicing by channel, F′ high Represents the high-level feature map after convolution module processing, F mid Represents the original mid-level feature map, F′ low Represents the low-level feature map after depth-wise separable convolution.

[0059] (2.4) Repeat steps (2.1) to (2.3), process the feature maps of different levels from top to bottom, complete the fusion of multi-scale features, and strengthen the semantic information of shallow feature maps, and finally obtain the extremely small scale feature f′ after feature fusion tiny , low-level features f′ low , intermediate feature f′ mid and high-level features f′ high Steps (2.1) to (2.3) are executed three times in total. Starting from the high-level features output by the final SPPF module of the backbone network, different feature levels are gradually integrated from top to bottom to obtain intermediate features, low-level features, and extremely small-scale features.

[0060] (3) The fused feature map is input into the refined feature enhancement module (RFEM), which eliminates position conflicts and redundant information based on the spatiotemporal attention mechanism and optimizes the details and positioning capabilities of deep features, such as Figure 3 As shown;

[0061] (3.1) The extremely small scale feature f′ after feature fusion tinyThe image is sent to the C3STR module based on SwinTransformer, and the spatial and semantic expression capabilities of the feature map are enhanced through the self-attention mechanism. SwinTransformer can focus on specific important areas while maintaining global features, which is suitable for the refined processing of low-level feature maps. The C3TR module used to process the lowest-level features is replaced by the C3STR module based on SwinTransformer, which outputs the extremely small-scale features f″ after feature enhancement. tiny .

[0062] (3.2) The deep separable convolution of the lowest level multi-level feature fusion module ML-FFM is used to extract the class-independent mask map of the dense area of ​​small objects; this mask map can indicate the approximate distribution area of ​​small objects in the feature map. First, a deep separable convolution with a convolution kernel size of 13×13 is used to expand the receptive field and capture the distribution information of small objects in high-resolution images; then, the features are mapped to a single channel through a 1×1 convolution, and the output of this channel will be used as the basis for generating the class-independent mask map. The single-channel feature map output by the 1×1 convolution is batch normalized to stabilize the feature distribution and ensure the consistency of the numerical range in subsequent processing; finally, the Sigmoid activation function is used to limit the value of the feature map to [0, 1].

[0063] (3.3) The feature obtained in step (3.1) is enhanced to obtain the smallest scale feature f″ tiny The low-level features f′ output by the multi-level feature fusion module ML-FFM low The splicing is then performed and the splicing result is sent to the Transformer-based C3TR module to extract contextual semantic feature information.

[0064] (3.4) The output of the Transformer-based C3TR module is used as the input feature f input It is sent to the lightweight convolutional attention module CBAM, which will derive the attention map along the two independent dimensions of channel and space, and then combine the attention map with the input feature f input The CBAM module can adaptively enhance feature expression, help the model focus on useful target objects, resist conflicting and redundant information generated by feature fusion, and effectively improve multi-scale detection performance.

[0065] (3.5) The output of the CBAM module is passed through a convolution module CBS to adjust the scale of the feature map to obtain the enhanced small-scale feature f″ small , f″ small Used to splice the feature map output by the next multi-level feature fusion module ML-FFM module.

[0066] (3.6) The intermediate feature f′ after fusion of the features obtained in step (2.4) mid And the enhanced small-scale feature f″ obtained in step (3.5) small As the input of step (3.3), repeat steps (3.3) to (3.5) and use the refined feature enhancement module RFEM to enhance the feature maps of different levels from bottom to top to obtain the feature maps of different scales after feature enhancement, which are the extremely small scale features f″ tiny , small-scale features f″ small , mesoscale features f″ mid and large-scale features f″ big .

[0067] (4) The enhanced feature map is sent to the Transformer-based prediction head for detection to complete the target classification and positioning of the UAV aerial image.

[0068] (4.1) For the extremely small scale feature f″ in step (3.6) tiny , the target is detected through the Transformer-based sparse prediction head Sparse-TPH; Sparse-TPH only detects the target feature blocks that exceed the threshold γ in the mask image obtained in step (3.2), and for non-target feature blocks, Sparse-TPH uses sparse convolution to reduce the computational waste on the background.

[0069] (4.2) For the feature maps {f″ small , f″ mid , f″ big}, detect the target through the Transformer-based prediction head TPH of the corresponding scale.

[0070] Obviously, those skilled in the art should understand that the various steps of the method for detecting a target from a drone's perspective based on feature fusion and detection head collaborative optimization of the above-mentioned embodiment of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than that here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims

1. A method for detecting unmanned aerial vehicle targets based on feature fusion and detection head collaborative optimization, characterized in that: The steps include: (1) Extract features from the input drone aerial images and obtain multi-scale feature maps using a backbone network consisting of a stacked convolutional module (CBS) and a residual module (C2f); (2) The extracted multi-scale features are sent to the multi-level feature fusion module ML-FFM to fuse the high, medium and low-level features from deep to shallow layers to obtain the fused feature map; (3) The fused feature map is input into the refined feature enhancement module RFEM, which eliminates position conflicts and redundant information based on the spatiotemporal attention mechanism, optimizes the details and positioning ability of the deep features, and obtains the enhanced feature map; (4) The enhanced feature map is sent to the Transformer-based prediction head for detection to complete the target classification and positioning of the drone aerial image.

2. The method for detecting unmanned aerial vehicle targets based on feature fusion and detection head collaborative optimization according to claim 1, characterized in that: The implementation process of step (2) includes the following steps: (2.1) Convolve the high-level feature map to extract high-level semantic features, normalize the convolved features, and then apply the SiLU activation function to obtain the processed high-level feature map; (2.2) Performing depth-wise separable convolution on the low-level feature map to obtain a processed low-level feature map; (2.3) Concatenate the upsampled high-level feature map, the original mid-level feature map, and the low-level feature map after depth-separable convolution, and extract the fused features through the residual module C2f; (2.4) Repeat steps (2.1) to (2.3), process the feature maps of different levels from top to bottom, complete the fusion of multi-scale features, and strengthen the semantic information of shallow feature maps, and finally obtain the extremely small scale features, low-level features, medium-level features and high-level features after feature fusion.

3. The method for detecting unmanned aerial vehicle targets based on feature fusion and detection head collaborative optimization according to claim 1, characterized in that: The specific implementation steps of step (3) are as follows: (3.1) The extremely small-scale features obtained in step (2.4) are fed into the C3STR module based on SwinTransformer, and the spatial and semantic expression capabilities of the feature map are enhanced through the self-attention mechanism, and the extremely small-scale features after feature enhancement are output; (3.2) Using the deep separable convolution of the lowest level multi-level feature fusion module ML-FFM to extract class-independent mask maps of small object dense areas; (3.3) The extremely small-scale features after feature enhancement obtained in step (3.1) are concatenated with the low-level features output by the multi-level feature fusion module ML-FFM, and then the concatenated results are sent to the Transformer-based C3TR module to extract contextual semantic feature information; (3.4) The output of the Transformer-based C3TR module is fed into the lightweight convolutional attention module CBAM as input features. CBAM derives the attention map along the two independent dimensions of channel and space, and then multiplies the attention map with the input features. (3.5) The output of the CBAM module is passed through a convolutional module CBS to adjust the scale of the feature map to obtain enhanced small-scale features; (3.6) Using the intermediate features after feature fusion obtained in step (2.4) and the enhanced small-scale features obtained in step (3.5) as the input of step (3.3), repeat steps (3.3) to (3.5), and use the refined feature enhancement module RFEM to enhance the feature maps of different levels from bottom to top to obtain feature maps of different scales after feature enhancement, namely, extremely small-scale features, small-scale features, medium-scale features, and large-scale features.

4. The method for detecting unmanned aerial vehicle targets based on feature fusion and detection head collaborative optimization according to claim 3 is characterized in that: In (3), a depth-wise separable convolution is used to generate a class-independent mask map for densely populated areas of small objects. First, a depth-wise separable convolution with a convolution kernel size of 13×13 is used to expand the receptive field and capture the distribution information of small objects in high-resolution images. Then, a 1×1 convolution is used to map the features to a single channel. The output of this channel will be used as the basis for generating a class-independent mask map. The single-channel feature map output by the 1×1 convolution is batch normalized to stabilize the feature distribution. Finally, a Sigmoid activation function is used to limit the value of the feature map to between [0,1].

5. The method for detecting unmanned aerial vehicle targets based on feature fusion and detection head collaborative optimization according to claim 1, characterized in that: The above-mentioned detection is performed by the Transformer-based prediction head to complete the target classification and positioning of the drone aerial image. The specific steps are as follows: (4.1) For the extremely small-scale features in step (3.6), the target is detected through the Transformer-based sparse prediction head Sparse-TPH; Sparse-TPH only detects the target feature blocks that exceed the threshold in the mask image obtained in step (3.2), and for non-target feature blocks, Sparse-TPH uses sparse convolution to reduce the computational waste on the background; (4.2) For feature maps of small, medium, and large scales, the target is detected through the Transformer-based prediction head TPH of the corresponding scale.

6. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for detecting unmanned aerial vehicle targets based on feature fusion and collaborative optimization of detection heads as described in any one of claims 1 to 5 is implemented.

7. A computer-readable storage medium storing a computer program for executing the unmanned aerial vehicle target detection method based on feature fusion and detection head collaborative optimization as described in any one of claims 1 to 5.

Images