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

Abstract

The application discloses a UAV target detection method based on feature fusion and detection head cooperative optimization, uses a backbone network stacked by a convolution module CBS and a residual module C2f to obtain multi-scale feature maps, sequentially sends the feature maps of various scales to a multi-level feature fusion module (ML-FFM), performs feature fusion from a deep layer to a shallow layer, and enhances semantic information of the shallow layer features; inputs the fused feature maps into a refined feature enhancement module (RFEM), eliminates position conflicts and redundant information based on a space-time attention mechanism, optimizes details and positioning capability of the deep layer features; sends the enhanced feature maps into a prediction head based on a Transformer for detection, and completes target classification and positioning of UAV aerial images. The application solves the problems of large scale change of the UAV aerial images, and small target proportion and dense distribution, thereby improving the precision of target detection in the UAV aerial images. The modules proposed in the method can significantly reduce the calculation amount, and are suitable for real-time detection scenes of the UAV.

Description

TECHNICAL FIELD

[0001] The application relates to a UAV target detection method based on feature fusion and detection head cooperative optimization, in particular to target detection under the condition of high small target proportion and dense distribution in the UAV visual angle, and belongs to the technical field of target detection. BACKGROUND

[0002] Target detection, as an important research direction of computer vision and digital image processing, is widely used in intelligent auxiliary driving, industrial detection and other fields. Through target detection technology, computer vision algorithms can greatly reduce human consumption and provide more accurate services. Target detection uses basic target classification network to further realize automatic identification of objects in pictures or videos, and has become a research hotspot in the field of computer vision and artificial intelligence. In the application of UAV, target detection endows UAV with the ability of autonomous perception, analysis and decision-making, and makes it show a wide application prospect in the fields of traffic monitoring, power inspection, water inspection and disaster rescue. UAV has unique aerial perspective and high mobility, and can realize efficient data collection, especially in traffic monitoring, which shows the advantages of fast response and wide field of view. However, most of the current target detection research focuses on ground video monitoring, and the target detection of UAV visual angle image has not been fully explored, and the accuracy of existing algorithms on UAV image is generally low.

[0003] Target detection under the visual angle of UAV faces many challenges. First, the flexible flight height and variable visual angle of UAV lead to significant changes in the distance of the photographed object, and the scale of the target in the image is extremely uneven. In addition, the wide field of view of UAV covers a large number of targets and a large amount of background, resulting in the clustering distribution of targets in the image: the target is dense in some areas, there is a shielding problem, and the target is sparse or even no target in some areas, which increases the waste of computing resources. Compared with ground visual angle images, the number of foreground targets in UAV visual angle is larger, but the proportion of targets in the image is small, which leads to unbalanced positive and negative samples, and the model is easily disturbed by background noise, making it difficult to accurately focus on the target. Further, since the image resolution of the UAV dataset is usually high, processing high-resolution images requires a large amount of video memory and computing resources, and the chip of the UAV platform is limited by low power consumption, which is difficult to bear high computing load. In addition, real-time processing of high-resolution images requires low delay, so it is necessary to balance the computing load and detection accuracy. SUMMARY

[0004] Invention purposes: In view of the problems and deficiencies of the prior art, it is necessary to fully extract the feature information of the unmanned aerial vehicle aerial image, enhance the features of small targets through feature fusion, and then improve the accuracy of unmanned aerial vehicle carrier target detection. However, the proportion of small targets in the unmanned aerial vehicle aerial image is large, and the target is dense in a small part of the area. If the whole high-resolution image is processed, a large amount of computing resources will be wasted. In order to overcome this problem, the present application provides a kind of unmanned aerial vehicle visual angle target detection method based on feature fusion and detection head cooperative optimization. By extracting the multi-scale features of the unmanned aerial vehicle aerial image, and effectively fusing the context information of different levels of features, a sparse detection head is introduced to accurately locate the small target, which not only greatly improves the accuracy and efficiency of the unmanned aerial vehicle visual angle target detection, but also reduces the calculation amount of the model, which has important significance for real-time target detection of unmanned aerial vehicles.

[0005] Technical scheme: A kind of unmanned aerial vehicle target detection method based on feature fusion and detection head cooperative optimization, comprising the following steps:

[0006] (1) feature extraction is carried out on the input unmanned aerial vehicle aerial image, and multi-scale feature maps are obtained by using the backbone network stacked by convolution module CBS and residual module C2f;

[0007] (2) the multi-scale features extracted are sent to the multi-level feature fusion module ML-FFM, the high, medium and low layer features are fused from deep to shallow, the fused feature maps are obtained, and the semantic features of the shallow feature maps are strengthened;

[0008] (3) the fused feature maps are input into the refined feature enhancement module (RFEM), the position conflict and redundant information are eliminated based on the space-time attention mechanism, the details and positioning ability of the deep features are optimized, and the enhanced feature maps are obtained;

[0009] (4) the enhanced feature maps are sent to the prediction head based on Transformer for detection, and the target classification and positioning of the unmanned aerial vehicle aerial image are completed.

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

[0011] (2.1) first, the high-level feature map is convolved, the high-level semantic feature is extracted, the normalized feature after convolution is stabilized, and then the SiLU activation function is applied to obtain the processed high-level feature map F′ high :

[0012]

[0013] Where F′ high represents the processed high-level feature map, and Fhigh denotes the original high-level feature map, denotes the up-sampling operation, denotes the SiLU activation function, denotes the batch normalization, denotes the convolution operation.

[0014] (2.2) Then, a depth separable convolution is performed on the low-level feature map, effectively reducing the amount of calculation while preserving important spatial information, to obtain the processed low-level feature map F' low :

[0015]

[0016] wherein F' low denotes the processed low-level feature map, F low denotes the original low-level feature map, denotes the depth convolution, denotes the point-wise convolution.

[0017] (2.3) Finally, the up-sampled high-level feature map, the original middle-level feature map, and the low-level feature map processed by the depth separable convolution are spliced, and the fused features are extracted through the residual module C2f to obtain F fusion :

[0018]

[0019] wherein F fusion denotes the output of the multi-level feature fusion module ML-FFM, denotes the residual module C2f, denotes the channel-wise splicing, F' high denotes the high-level feature map processed by the convolution module, F mid denotes the original middle-level feature map, F' low denotes the low-level feature map processed by the depth separable convolution.

[0020] (2.4) Repeat steps (2.1) to (2.3) to 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, finally obtain the smallest scale feature f' after feature fusion tiny , low-level feature f' low , middle-level feature f' mid , and high-level feature f' high .

[0021] The RFEM based on the spatio-temporal attention mechanism eliminates position conflicts and redundant information, optimizes the details and positioning ability of deep features, and the specific steps are as follows:

[0022] (3.1) The minimal scale feature f′ after feature fusion tiny The data is fed into the SwinTransformer-based C3STR module, which uses the self-attention mechanism to enhance the spatial and semantic expression capabilities of the feature map. SwinTransformer can focus on specific important areas while maintaining global features, making it suitable for fine-grained processing of low-level feature maps. The C3TR module used to process the lowest-level features is replaced by the SwinTransformer-based C3STR module, which outputs the enhanced very small-scale features f″ tiny .

[0023] (3.2) The deep separable convolution of the lowest-level multi-level feature fusion module ML-FFM is used to extract a class-independent mask map for densely populated areas 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 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 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 a consistent numerical range in subsequent processing. Finally, a sigmoid activation function is used to limit the value of the feature map to the range [0, 1].

[0024] (3.3) The minimum scale feature f″ obtained by step (3.1) is enhanced tiny The low-level features f′ output by the multi-level feature fusion module ML-FFM low The splicing is performed and then 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 derives the attention map along the two independent dimensions of channel and space, and then combines 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 convolutional module CBS to adjust the scale of the feature map to obtain the enhanced small-scale feature f″ small , f″ smallfor splicing with the feature map output by the next layer multi-level feature fusion module ML-FFM module.

[0027] (3.6) the intermediate features f' obtained after feature fusion in step (2.4) mid and the enhanced small-scale features f'' obtained in step (3.5) small As the input of step (3.3), steps (3.3) to (3.5) are repeated, and the fine feature enhancement module RFEM is used to enhance the feature maps of different levels from bottom to top, to obtain feature-enhanced feature maps of different scales, which are respectively the extremely small-scale features f'' tiny , the small-scale features f'' small , the medium-scale features f'' mid and the large-scale features f'' big .

[0028] The detection by the Transformer-based prediction head completes the target classification and positioning of the UAV aerial image, and the specific steps are as follows:

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

[0030] (4.2) For the three scale feature maps {f'' small , f'' mid , f'' big}, the target is detected by the corresponding scale Transformer-based prediction head TPH.

[0031] The convolution module CBS in the UAV perspective target detection method consists of a 3x3 convolution layer, a batch normalization layer and a SiLU activation function.

[0032] The residual module C2f in the UAV perspective target detection method consists of a 1x1 convolution layer, a plurality of 3x3 residual blocks Bottleneck and a 1x1 convolution layer. First, the input channels are divided into two parts by 1x1 convolution, Bottleneck blocks are stacked layer by layer to extract deep features, and finally, channel fusion is performed by 1x1 convolution to realize feature dimension reduction and fusion, effectively balancing detailed information and global semantic information, suitable for multi-scale feature extraction and efficient calculation.

[0033] The C3TR module in the UAV visual angle target detection method is composed of a multi-head self-attention layer and a multi-layer perception layer, and each sub-layer is provided with layer normalization, Dropout and residual connection.

[0034] The multi-level feature fusion module ML-FFM in the UAV visual angle target detection method fuses features of three levels of low, medium and high from top to bottom through multi-level feature fusion.

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

[0036] A computer device includes a memory, a processor and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to realize the UAV target detection method based on feature fusion and cooperative optimization of detection heads as described above.

[0037] A computer readable storage medium stores a computer program for executing the UAV target detection method based on feature fusion and cooperative optimization of detection heads as described above.

[0038] Advantages: Compared with the prior art, the present application has the following advantages:

[0039] (1) The multi-level feature fusion module ML-FFM can fuse feature maps of three sizes of low, medium and high, balance the spatial information of low-level features and the semantic information of high-level features, thereby enhancing the recognition and classification ability of target position, and helping the detection and fine positioning of small targets.

[0040] (2) The fine feature enhancement module RFEM uses modules such as SwinTransformer, C3TR and CBAM to refine the fused features at multiple levels by using self-attention, context semantics and channel-spatial attention mechanisms. SwinTransformer focuses on global semantic information in specific regions, the C3TR module extracts context semantic information, and the CBAM module adaptively removes redundant features, so that the network can more effectively capture key feature regions, helping the model focus on useful target features in complex detection tasks and improving detection accuracy.

[0041] (3) In the detection of extremely small scale targets, the sparse prediction head Sparse-TPH based on Transformer only detects target feature regions that exceed a threshold, and sparse convolution is used in other regions. This strategy significantly reduces the computational overhead of background regions, improves the computational efficiency of the network, and is suitable for processing high-resolution UAV aerial images. For small, medium and large scale feature maps, TPH only applies Transformer encoder blocks at the end of the network on low resolution feature maps, reducing the computational and memory costs while maintaining good feature expression ability. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 Fig. 1 is a network structure diagram of a UAV target detection method based on feature fusion and detection head cooperative optimization in specific embodiments;

[0043] Figure 2 Fig. 3 is a structure diagram of a multi-level feature fusion module ML-FFM in specific embodiments;

[0044] Figure 3 Fig. 4 is a structure diagram of a fine feature enhancement module RFEM in specific embodiments. DETAILED DESCRIPTION

[0045] The present application will be further illustrated below in conjunction with specific embodiments, which are intended to illustrate the present application but not to limit the scope of the present application. After reading the present application, those skilled in the art can make various equivalent modifications to the present application, which fall within the scope defined by the claims attached hereto.

[0046] It is known that in a certain urban traffic monitoring system, a UAV carries a high-definition camera for high-altitude cruising, and real-time shooting of traffic flow images of urban roads. These images may contain 4 types of traffic targets, namely small cars, buses, trucks and pedestrians. It is necessary to extract UAV image features through a model, distinguish foreground and background, accurately detect and classify various traffic targets, distinguish different traffic flow, and realize 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 cooperative optimization is given, and the specific implementation is as follows:

[0048] (1) Feature extraction is performed on the input UAV aerial image, and a backbone network stacked by a convolution module CBS and a residual module C2f is used to obtain multi-scale feature maps;

[0049] (2) The extracted multi-scale features are sent to the multi-level feature fusion module ML-FFM, and high, medium and low level feature fusion is performed from deep to shallow, strengthening the semantic features of shallow feature maps, as shown in Figure 2 ;

[0050] (2.1) First, the high-level feature map is convolved to extract high-level semantic features, and the normalized features are applied 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 up-sampling operation, represents the SiLU activation function, represents the batch normalization, represents the convolution operation.

[0053] (2.2) Then, the low-level feature map is depthwise separable convolution, which effectively reduces the amount of calculation while preserving important spatial information, and obtains the processed low-level feature map F′ low :

[0054]

[0055] where F′ low represents the processed low-level feature map, F lowdenotes the original low-level feature map, denotes the deep convolution, denotes the point-wise convolution.

[0056] (2.3) Finally, the up-sampled high-level feature map, the original middle-level feature map and the low-level feature map processed by the depth separable convolution are spliced, and the fused features are extracted by the residual module C2f to obtain F fusion :

[0057]

[0058] wherein F fusion denotes the output of the multi-level feature fusion module ML-FFM, denotes the residual module C2f, denotes the channel-wise splicing, F′ high denotes the high-level feature map processed by the convolution module, F mid denotes the original middle-level feature map, F′ low denotes the low-level feature map processed by the depth separable convolution.

[0059] (2.4) Repeat steps (2.1) to (2.3) to 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 smallest scale feature f′ tiny , low-level feature f′ low , middle-level feature f′ mid and high-level feature f′ high . Steps (2.1) to (2.3) are performed a total of 3 times. Starting from the high-level feature output by the SPPF module at the end of the backbone network, the fusion of different feature levels is performed from top to bottom step by step to obtain the middle-level feature, the low-level feature and the smallest scale feature.

[0060] (3) The fused feature map is input into the refined feature enhancement module (RFEM), and based on the spatio-temporal attention mechanism, the position conflict and redundant information are eliminated to optimize the details and positioning ability of deep features, as shown in Figure 3 .

[0061] (3.1) The smallest scale feature f′ tinyThe data is fed into the SwinTransformer-based C3STR module, which uses the self-attention mechanism to enhance the spatial and semantic expression capabilities of the feature map. SwinTransformer can focus on specific important areas while maintaining global features, making it suitable for fine-grained processing of low-level feature maps. The C3TR module used to process the lowest-level features is replaced by the SwinTransformer-based C3STR module, which outputs the enhanced very small-scale features f″ tiny .

[0062] (3.2) The deep separable convolution of the lowest-level multi-level feature fusion module ML-FFM is used to extract a class-independent mask map for densely populated areas 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 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 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 a consistent numerical range in subsequent processing. Finally, a sigmoid activation function is used to limit the value of the feature map to the range [0, 1].

[0063] (3.3) The minimum scale feature f″ obtained by step (3.1) is enhanced tiny The low-level features f′ output by the multi-level feature fusion module ML-FFM low The splicing is performed and then 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 derives the attention map along the two independent dimensions of channel and space, and then combines 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 convolutional 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 layer of multi-level feature fusion module ML-FFM module.

[0066] (3.6) the intermediate-level features f' obtained after feature fusion in step (2.4) mid and the enhanced small-scale features f" obtained in step (3.5) small As the input of step (3.3), steps (3.3) to (3.5) are repeated to perform enhancement processing on feature maps of different levels from bottom to top using the refinement feature enhancement module RFEM, to obtain feature-enhanced feature maps of different scales, which are respectively the extremely small-scale features f" tiny , the small-scale features f" small , the medium-scale features f" mid and the large-scale features f" big .

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

[0068] (4.1) For the extremely small-scale features f" tiny obtained in step (3.6), the Transformer-based sparse prediction head Sparse-TPH is used to detect targets; Sparse-TPH only detects target feature blocks in the mask map obtained in step (3.2) that exceed the threshold γ, and for non-target feature blocks, Sparse-TPH uses sparse convolution to reduce the waste of calculation on the background.

[0069] (4.2) For the small-scale, medium-scale and large-scale feature maps {f" small , f" mid , f" big}, the corresponding scale Transformer-based prediction head TPH is used to detect targets.

[0070] Obviously, those skilled in the art should understand that the above-mentioned UAV visual angle target detection method based on feature fusion and detection head collaborative optimization of the embodiments of the present application can be realized by general computing devices, which can be concentrated on a single computing device or distributed on a network composed of multiple computing devices, and optionally, they can be realized by program codes executable by computing devices, so that they can be stored in storage devices and executed by computing devices, and in some cases, the steps shown or described can be executed in different order, or they can be manufactured into individual integrated circuit modules, or multiple modules or steps thereof can be manufactured into a single integrated circuit module. Thus, the embodiments of the present application are not limited to any specific combination of hardware and software.

Claims

1. A UAV target detection method based on feature fusion and detection head collaborative optimization, characterized in that: The steps include: (1) Feature extraction is performed on the input drone aerial image, and a multi-scale feature map is obtained using a backbone network consisting of a convolutional block (CBS) and a residual block (C2f). (2) The extracted multi-scale features are sent to the multi-level feature fusion module ML-FFM, and the high, medium and low-level features are fused from deep to shallow layers to obtain a fused feature map. The fused feature map includes extremely small-scale features, low-level features, medium-level features and high-level features; (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 fed into the Transformer-based prediction head for detection to complete the target classification and localization of the UAV aerial image; The specific implementation steps of step (3) are as follows: (3.1) The extremely small-scale features 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 enhanced extremely small-scale features are output; (3.2) Utilize the depthwise separable convolution of the lowest-level multi-level feature fusion module ML-FFM to extract class-independent mask maps of densely populated areas of small objects; (3.3) The enhanced microscale features obtained in step (3.1) are concatenated with the low-level features output by the multi-level feature fusion module ML-FFM. The concatenated results are then fed into the Transformer-based C3TR module to extract contextual semantic feature information. (3.4) The output of the Transformer-based C3TR module is fed into a lightweight convolutional attention module (CBAM) as input features. CBAM derives attention maps along two independent dimensions, channel and spatial, and then multiplies the attention maps with the input features. (3.5) The output of the CBAM module is passed through a convolutional block (CBS) to adjust the scale of the feature map and obtain enhanced small-scale features. (3.6) Using the intermediate features after feature fusion 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. The feature maps of different scales after feature enhancement are obtained, namely, very small-scale features, small-scale features, medium-scale features, and large-scale features. The Transformer-based prediction head performs detection to complete target classification and positioning of drone aerial images. The specific steps are as follows: (4.1) For the extremely small-scale features in step (3.6), the target is detected by the Transformer-based sparse prediction head Sparse-TPH; Sparse-TPH only detects the features that exceed the threshold in the mask image obtained in step (3.2). The target feature blocks are detected, while 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.

2. The UAV target detection method based on feature fusion and detection head collaborative optimization according to claim 1 is 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 the 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 depthwise 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 enhance 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 UAV target detection method based on feature fusion and detection head collaborative optimization according to claim 1 is characterized in that: In the step (3), depthwise separable convolution is used to generate a class-independent mask map for dense areas of small objects. First, a depthwise separable convolution with a convolution kernel size of 13×13 is used to expand the receptive field to 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. 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 [0, 1].

4. 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 a target of an unmanned aerial vehicle (UAV) based on feature fusion and detection head collaborative optimization as described in any one of claims 1 to 3 is implemented.

5. A computer-readable storage medium storing a computer program for executing the UAV target detection method based on feature fusion and detection head collaborative optimization as described in any one of claims 1 to 3.

Images