An urban remote sensing embedded multi-target detection and positioning method and system

Abstract

This invention discloses an embedded multi-target detection and positioning method for urban remote sensing. After acquiring an input image, a specially structured high-speed buffer controller directly exchanges data with a GPU. Feature extraction and fusion are performed to obtain a three-scale feature map. The target detection results from the detection head are then combined with multi-sensor information in a multi-threaded calculation to locate the target. The fused positioning results are then output by a GStreamer pipeline receiver component. The invention also discloses a system for implementing the method. This invention provides a complete solution for urban remote sensing target detection and positioning tasks in an embedded device, characterized by low computational effort, high detection accuracy, and strong flexibility.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing and computer vision technology, and in particular to an urban remote sensing embedded multi-target detection and positioning method and system. Background Art

[0002] Computer vision has been widely used in various scenarios, enabling the perception, analysis, and decision-making of information such as equipment, environments, and people. Object detection, as an application of computer vision, can automatically detect and identify objects in images or long-duration videos, reducing the need for human intervention and improving safety and security. With the acceleration of urbanization and the continuous growth in the size and number of cities, object detection technology plays a vital role in accurately extracting and analyzing urban feature information, monitoring urban environmental conditions, and responding to natural disasters. Embedded object detection technology not only enables the rapid and accurate extraction of urban feature information, providing effective data support for urban planning and management, but also enables the rapid acquisition of information about affected areas after a disaster, providing data support for post-disaster rescue and reconstruction.

[0003] Thanks to the rapid development of embedded hardware and deep learning networks, object detection tasks can now be run on some embedded devices. However, ensuring accuracy and real-time performance remains challenging and presents numerous challenges. First, the hardware resources of embedded systems are relatively limited, necessitating optimization to improve the efficiency of embedded object detection algorithms. Second, embedded object detection tasks often require high real-time performance in scenarios such as traffic safety and security monitoring. Therefore, algorithms require real-time optimization to ensure they can complete object detection and recognition tasks in a relatively short time. Furthermore, embedded object detection tasks often operate in complex and changing environments. For example, the target detection process in drone remote sensing involves varying lighting, weather, and object angles, requiring robust models. Finally, to achieve real-time performance, embedded object detection algorithms typically require relatively low complexity. Therefore, it is crucial to minimize algorithm complexity while ensuring effective detection. Therefore, it is crucial to research how to leverage the advantages of embedded devices, such as privacy, widespread application, and cost savings, to make embedded object detection tasks more efficient, real-time, and robust. Summary of the Invention

[0004] This application proposes an urban remote sensing embedded multi-target detection and positioning method and system to enhance the key coupling of embedded detection system modules and solve the problems of missed detection, wrong detection and poor real-time performance in remote sensing target detection.

[0005] On the one hand, the embodiment of the present application proposes an urban remote sensing embedded multi-target detection and positioning method, the specific steps are as follows:

[0006] Acquire image data, complete labeling, cleaning and segmentation, perform data analysis and offline data enhancement, and obtain a data set;

[0007] Build a deep learning network, including a structured reparameterized feature extraction network and an effective residual path aggregation network;

[0008] Using the data set to perform hyperparameter adjustment and server training on the deep learning network until the network parameters no longer converge or the training time reaches a preset number of rounds, and obtain pre-trained weights;

[0009] Replace the activation function in the original network with a trainable truncated activation function, load the pre-trained weights to fine-tune the network, and retain the final network structure file and weight parameter file;

[0010] Configure target detection logic in embedded devices, perform multi-sensor integration, obtain multi-sensor fusion information, and deploy network structure files and weight parameter files;

[0011] Build a GStreamer multimedia pipeline and connect the target detection logic and sensor module as components to the pipeline;

[0012] Embedded devices acquire video streams and posture information in real time, open up multiple threads, and output multi-sensor fusion information to target detection logic;

[0013] The inference process is based on the multi-threaded scheduling method of CPU, GPU and deep learning hardware accelerator. The image data is directly transferred from the high-speed cache controller to the GPU to perform prediction and output the positioning results.

[0014] Preferably, the target detection logic uses a deep learning network, a parallel computing library, and a deep learning hardware accelerator, and includes the following steps:

[0015] Set compilation configuration parameters;

[0016] The model weights are saturated and quantized using TensorRT, and a serialized model is generated through the constructor;

[0017] During the inference phase, the inference algorithm loads the serialized quantized model, uses the parallel computing library, and calls the deep learning hardware accelerator to accelerate the model.

[0018] Preferably, the cache controller includes a cache data storage structure, the GPU and the secondary cache controller directly exchange data, the CacheSRAM adopts a four-way group-associative mapping structure in which components are directly connected and fully connected within the group, each storage unit is a cache block, each cache block contains 8 small cache blocks, each small cache block is 4 bytes and each small cache block has a read and write enable signal; read and write operations on each small cache block can be realized according to the enable signal, thereby improving the utilization rate of the CacheSRAM.

[0019] Preferably, the multi-thread scheduling method schedules the multi-core CPU, GPU and DLA of the embedded device so that after one thread completes the GPU calculation, the second thread immediately starts the graphics calculation task.

[0020] Preferably, the multi-sensor integration includes a posture sensor and a GPS module, and data integration is performed through a specific coding and decoding protocol. The posture sensor integrates a gyroscope, an accelerometer and a geomagnetic sensor.

[0021] Preferably, the multi-sensor fusion information, GPS module and attitude sensor are connected to the embedded device through the UART serial port, and the integrated data binary result is output to the target detection logic according to the set encoding and decoding logic. During the inference process, multi-threaded fusion calculation is opened with the detection head feature map information, instead of waiting for the detection head to output the result and then serially reading the multi-sensor fusion information.

[0022] Preferably, the trainable truncated activation function is characterized by setting a trainable parameter α, limiting the ReLu activation function x-axis positive half axis [α, +∞) range to α, (-∞, 0) range to 1 / α, thereby reducing outliers in the non-saturated quantization process.

[0023] Preferably, the structure reparameterizes the feature extraction network, includes a residual multi-branch structure, and the module structure is reparameterized during the inference process, and feature extraction is performed only through 3×3 convolution.

[0024] Preferably, the effective residual path aggregation network includes a CSP residual structure, and uses a channel attention module to adaptively adjust multi-level residual information, giving channels containing effective information greater weights, and enhancing feature fusion capabilities.

[0025] On the other hand, the present application also proposes an urban remote sensing embedded multi-target detection and positioning system for implementing the method described in any embodiment of the present application, including:

[0026] The server side is used for deep learning network training, obtaining pre-trained weights and performing truncated activation quantization, and obtaining network structure files and weight parameter files;

[0027] The embedded side includes sensor modules, target detection logic, and embedded devices. The sensor modules and target detection logic are mounted as components in the GStreamer pipeline.

[0028] The sensor module is used to obtain video stream and posture information in real time and output it to the target detection logic;

[0029] The target detection logic is used to perform prediction and output positioning results;

[0030] The embedded device includes a CPU, a GPU and a high-speed cache controller, and the high-speed cache controller is used to execute a multi-threaded scheduling mode.

[0031] The beneficial effects of the technical solution provided by this application include at least:

[0032] Targets in urban remote sensing scenarios have scale diversity and perspective diversity. To address the shortcomings of existing general target detection methods and system limitations, the present invention proposes a real-time multi-target detection system based on embedded devices in complex backgrounds. The system is characterized by high efficiency, high flexibility and easy implementation. The system detection process is based on the GStreamer pipeline design, including target detection logic and multi-sensor integration. Each module is highly integrated. At the same time, an algorithm is designed to enhance the key features of multi-target and small target images, realizing comprehensive and accurate detection of multi-scale targets in urban remote sensing backgrounds. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments.

[0034] Figure 1 This is a flow chart of an embodiment of the urban remote sensing multi-target real-time detection and positioning method of this application;

[0035] Figure 2 This is the workflow diagram of the target detection logic;

[0036] Figure 3 The present invention provides an embedded system structure for real-time detection and positioning of multiple urban remote sensing targets;

[0037] Figure 4 is the embedded target detection logic of the system of the present invention;

[0038] Figure 5 It is the multi-sensor integration and multi-sensor information fusion module of the present invention;

[0039] Figure 6 It is the server training structure of the present invention;

[0040] Figure 7 It is the residual multi-branch structure of the present invention;

[0041] Figure 8 It is a CSP residual structure proposed in the present invention;

[0042] Figure 9 This is the multi-thread scheduling method in the present invention. DETAILED DESCRIPTION

[0043] The following will further describe the present invention based on this technical solution in combination with specific examples and drawings, and provide specific implementation methods and operation process examples.

[0044] To achieve the above functions, the present invention includes: a server, a sensor module, a target detection logic and an embedded device;

[0045] Figure 1 This is a flow chart of an embodiment of the urban remote sensing multi-target real-time detection and positioning method of this application. An urban remote sensing embedded multi-target detection and positioning method, the specific steps are as follows:

[0046] Step 110: Acquire image data, complete labeling, cleaning, and segmentation, perform data analysis and offline data enhancement, and obtain a data set;

[0047] Step 120: Building a deep learning network, including a structure reparameterization feature extraction network and an effective residual path aggregation network;

[0048] Step 130: Using the data set, perform hyperparameter adjustment and server training on the deep learning network until the network parameters no longer converge or the training time reaches a preset number of rounds, and obtain pre-trained weights;

[0049] Step 140: Replace the activation function in the original network with a trainable truncated activation function, load the pre-trained weights to fine-tune the network, and retain the final network structure file and weight parameter file;

[0050] The trainable truncated activation function is characterized by setting a trainable parameter α, limiting the ReLu activation function x-axis positive half axis [α, +∞) range to α, (-∞, 0) range to 1 / α, and reducing outliers in the non-saturated quantization process.

[0051] Step 150: Configure the target detection logic in the embedded device and read the multi-sensor fusion information, and deploy the network structure file and weight parameter file in it;

[0052] The multi-sensor integration is characterized by including a posture sensor and a GPS module, performing data integration through a specific coding and decoding protocol, and the posture sensor integrates a gyroscope, an accelerometer and a geomagnetic sensor.

[0053] Step 160: Building a GStreamer multimedia pipeline, connecting the target detection logic and multi-sensor integration as components to the pipeline;

[0054] Step 170: The embedded device obtains the video stream and posture information in real time, opens multiple threads, and outputs the multi-sensor fusion information to the target detection logic;

[0055] The sensor fusion information, GPS module and attitude sensor are connected to the embedded device through the UART serial port. The integrated data is sent to the embedded device for dynamic solution and filtering according to the set encoding and decoding logic, and the binary result is output to the target detection logic. During the inference process, multi-threaded fusion calculation is opened with the detection head feature map information, instead of waiting for the detection head to output the result and then serially reading the sensor fusion information.

[0056] Step 180: The inference process is based on the multi-threaded scheduling mode of the CPU, GPU and deep learning hardware accelerator. The image data is directly transmitted from the high-speed cache controller to the GPU to perform prediction and output the positioning result.

[0057] The high-speed cache controller includes a cache data storage structure. The GPU and the secondary high-speed cache controller directly exchange data. The Cache SRAM adopts a four-way group-associative mapping structure with direct component connection and full connectivity within the group. Each storage unit is a cache block, and each cache block contains 8 small cache blocks. Each small cache block is 4 bytes and each small cache block has a read and write enable signal. Read and write operations can be performed on each small cache block according to the enable signal, thereby improving the utilization rate of the Cache SRAM.

[0058] Figure 2 This is the workflow diagram of the target detection logic.

[0059] The target detection logic includes a deep learning network, a parallel computing library, and a deep learning hardware accelerator. The implementation steps are as follows:

[0060] Step 210: Set TensorRT compilation configuration parameters;

[0061] Step 220: The model weights are saturated quantized using TensorRT, and a serialized model is generated through the constructor;

[0062] Step 230: In the inference phase, the inference algorithm loads the serialized model, uses the parallel computing library, and calls the deep learning hardware accelerator to accelerate the model.

[0063] The structure reparameterizes the feature extraction network, including a residual multi-branch structure. The module structure is reparameterized during the inference process, and feature extraction is performed only through 3×3 convolution.

[0064] The effective residual path aggregation network includes a CSP residual structure and uses a channel attention module to adaptively adjust multi-level residual information, giving channels containing effective information greater weights and enhancing feature fusion capabilities.

[0065] Figure 3 The present invention provides an embedded system structure for real-time detection and positioning of multiple targets in urban remote sensing, with the purpose of achieving real-time detection and positioning of targets in urban remote sensing scenarios. Specifically, an embedded system for real-time detection and positioning of multiple targets in urban remote sensing includes:

[0066] The server side is used for deep learning network training, obtaining pre-trained weights and performing truncated activation quantization, and obtaining network structure files and weight parameter files;

[0067] The embedded side includes sensor modules, target detection logic, and embedded devices. The sensor modules and target detection logic are mounted as components in the GStreamer pipeline.

[0068] Furthermore, the sensor module is integrated with multiple sensors to obtain at least external air pressure, longitude and latitude information and perform information fusion. After encoding and dynamic calculation, the multi-sensor fusion information is obtained and output to the target detection logic for positioning;

[0069] Furthermore, the target detection logic includes at least a parallel computing library, a deep learning network, and a deep learning accelerator, which are mainly used to perform target detection and simultaneously fuse the multi-sensor fusion information to calculate the positioning result;

[0070] Furthermore, the embedded device includes a CPU, a GPU and a high-speed cache controller, which on the one hand receives multi-sensor fusion information and deep learning network inference results of target detection logic, and on the other hand controls the sensor module and target detection logic and performs multi-threaded scheduling on them.

[0071] like Figure 4 As shown in the figure, embedded target detection logic configures TensorRT compilation configuration parameters, obtains and compiles network structure files and weight parameter files, obtains saturation quantization to obtain a serialized model, generates a serialized model through the constructor, and executes the inference algorithm. TensorRT enters the execution phase and outputs the serialized model to the deep learning hardware accelerator for acceleration;

[0072] like Figure 5As shown in the figure, the multi-sensor integration in the sensor module includes a GPS module and an attitude sensor. After dynamic solution and filtering, longitude and latitude information, acceleration information, geomagnetic information, angle information and atmospheric pressure are obtained, and fused to obtain multi-sensor fusion information.

[0073] Furthermore, the multi-sensor fusion information is output to the embedded target detection logic, and the high-speed cache controller in the embedded device performs multi-threaded scheduling on the CPU, GPU and deep learning hardware accelerator to output the positioning results.

[0074] like Figure 6 As shown in the figure, the server-side structure is used for model training. The specific implementation steps are as follows:

[0075] Install Anaconda, configure the Python environment, install necessary deep learning libraries, install the Compute Unified Device Architecture (CUDA) and the Deep Neural Network library (cuDNN) to implement GPU multi-threaded computing.

[0076] Furthermore, images were obtained through personal collection and public datasets, and manually annotated in XML format using the labelImg tool to obtain the original images and apply data augmentation strategies, including color transformation, random geometric transformation, and Gaussian random noise addition. Python was used to extract categories based on the targets contained in the visible light images and the corresponding label files, obtaining the required categories and dividing them into training, validation, and test sets. The matplotlib library was also used for data analysis. The k-means algorithm was used to cluster the target size and aspect ratio of the training set data to obtain the anchor box size and aspect ratio required for network training. The results were compared with the Python visualization analysis results to verify the reliability and scientificity of the clustering results.

[0077] Furthermore, the k-means algorithm has the following specific steps:

[0078] Randomly select K points c1, c2, ..., c from the data set k As the center of the initial cluster, the center point is:

[0079] C={c1,c2,…,c k}

[0080] For each sample in the data set, calculate the distance from each sample to each cluster center point. The sample with the smallest distance to the cluster center point is divided into the class corresponding to the cluster center.

[0081] For each category i, recalculate the cluster center of the category Repeat the operation until the preset number of iterations is reached and the clustering result is obtained.

[0082] Furthermore, hyperparameter configuration is performed and an optimization strategy is formulated, and the training set and validation set are used for model training to obtain pre-training weights; the training weights are then used to evaluate the test set, and the evaluation results are saved.

[0083] Figure 7 The residual multi-branch structure of the present invention reduces the memory access cost of the reasoning process. Figure 7 The residual multi-branch structure shown is used to construct a structural reparameterized feature extraction network; in the training phase, the residual multi-branch structure includes two convolutional layers composed of residual structures, and the input and output are connected by identity mapping. In the inference phase, it is transformed into two directly connected convolutional layers through the structural reparameterization idea, decoupling the training phase and the inference phase, and realizing the reparameterization of the network structure in the inference phase.

[0084] Figure 8 This is a CSP residual structure proposed in this paper, which solves the problem of unequal contribution of feature information and enhances the ability to detect small targets in urban remote sensing. Figure 8 As shown in the figure, the CSP residual structure is used to construct an effective residual path aggregation network; the CSP residual structure consists of two channels in parallel, the input feature map first passes through the convolution layer to extract features, and then is split into x0 and x1 according to the channel. x0 passes through multiple convolution layers containing identity mapping to extract features, and x1 is not processed in any way. After that, the feature maps of the intermediate feature map of x0 and x1 are spliced ​​to fully integrate the multi-level feature information. After passing through the channel attention module, effective features are selected, and then after passing through a convolution layer, the feature map is output.

[0085] Figure 9 This is the multi-threaded scheduling method in the present invention, where DLA stands for Deep Learning Accelerator, i.e., deep learning hardware accelerator. The present invention will use DLA and deep learning hardware accelerator interchangeably. Figure 9 As shown in the figure, it is a multi-threaded scheduling method that schedules the multi-core CPU, GPU and deep learning hardware accelerator of embedded heterogeneous devices. After one thread completes the GPU or DLA calculation, the second thread immediately starts the next graphics calculation task, avoiding the waiting time of image preprocessing.

[0086] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] Therefore, the present application also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present application.

[0088] Furthermore, the present application also proposes an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in any embodiment of the present application when executing the computer program.

[0089] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0090] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0092] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-permanent storage in a computer-readable medium, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0093] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0094] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. An urban remote sensing embedded multi-target detection and positioning method, characterized in that: The specific steps are as follows: Acquire image data, complete labeling, cleaning and segmentation, perform data analysis and offline data enhancement, and obtain a data set; Build a deep learning network, including a structured reparameterized feature extraction network and an effective residual path aggregation network; Using the data set to perform hyperparameter adjustment and server training on the deep learning network until the network parameters no longer converge or the training time reaches a preset number of rounds, and obtain pre-trained weights; Replace the activation function in the original network with a trainable truncated activation function, load the pre-trained weights to fine-tune the network, and retain the final network structure file and weight parameter file; Configure target detection logic in embedded devices, perform multi-sensor integration, obtain multi-sensor fusion information, and deploy network structure files and weight parameter files; Build a GStreamer multimedia pipeline and connect the target detection logic and sensor module as components to the pipeline; Embedded devices acquire video streams and posture information in real time, open up multiple threads, and output multi-sensor fusion information to target detection logic; The inference process is based on the multi-threaded scheduling method of CPU, GPU and deep learning hardware accelerator. The image data is directly transferred from the high-speed cache controller to the GPU to perform prediction and output the positioning results.

2. The urban remote sensing embedded multi-target detection and positioning method according to claim 1, wherein the target detection logic uses a deep learning network model, a parallel computing library, and a deep learning hardware accelerator, and is characterized by: Set compilation configuration parameters; The model weights are saturated and quantized using TensorRT, and a serialized model is generated through the constructor; During the inference phase, the inference algorithm loads the serialized quantized model, uses the parallel computing library, and calls the deep learning hardware accelerator to accelerate the model.

3. The urban remote sensing embedded multi-target detection and positioning method according to claim 1, characterized in that: The high-speed cache controller includes a cache data storage structure. The GPU and the secondary high-speed cache controller directly exchange data. The Cache SRAM adopts a four-way group-associative mapping structure with direct component connection and full connectivity within the group. Each storage unit is a cache block, and each cache block contains 8 small cache blocks. Each small cache block is 4 bytes and each small cache block has a read and write enable signal. Read and write operations can be performed on each small cache block according to the enable signal, thereby improving the utilization rate of the Cache SRAM.

4. The urban remote sensing embedded multi-target detection and positioning method according to claim 1, characterized in that: The multi-thread scheduling method schedules the multi-core CPU, GPU and DLA of the embedded device so that after one thread completes the GPU calculation, the second thread immediately starts the graphics calculation task.

5. The urban remote sensing embedded multi-target detection and positioning method according to claim 1, characterized in that: The multi-sensor integration includes a posture sensor and a GPS module, and performs data integration through a specific coding and decoding protocol. The posture sensor integrates a gyroscope, an accelerometer, and a geomagnetic sensor.

6. The urban remote sensing embedded multi-target detection and positioning method according to claim 1, characterized in that: The multi-sensor fusion information, GPS module and attitude sensor are connected to the embedded device through the UART serial port, and the integrated data binary results are output to the target detection logic according to the set encoding and decoding logic. During the inference process, multi-threaded fusion calculations are opened with the detection head feature map information, instead of waiting for the detection head to output the results and then serially reading the multi-sensor fusion information.

7. The urban remote sensing embedded multi-target detection and positioning method according to claim 1, characterized in that: The trainable truncated activation function is characterized by setting a trainable parameter α, limiting the ReLu activation function x-axis positive half axis [α, +∞) range to α, (-∞, 0) range to 1 / α, and reducing outliers in the non-saturated quantization process.

8. The urban remote sensing embedded multi-target detection and positioning method according to claim 1, characterized in that: The structure reparameterizes the feature extraction network, including a residual multi-branch structure. The module structure is reparameterized during the inference process, and feature extraction is performed only through 3×3 convolution.

9. The urban remote sensing embedded multi-target detection and positioning method according to claim 1, characterized in that: The effective residual path aggregation network includes a CSP residual structure and uses a channel attention module to adaptively adjust multi-level residual information, giving channels containing effective information greater weights and enhancing feature fusion capabilities.

10. An urban remote sensing embedded multi-target detection and positioning system, used to implement the urban remote sensing embedded multi-target detection and positioning method according to any one of claims 1 to 9, characterized in that: include: The server side is used for deep learning network training, obtaining pre-trained weights and performing truncated activation quantization, and obtaining network structure files and weight parameter files; The embedded side includes sensor modules, target detection logic, and embedded devices. The sensor modules and target detection logic are mounted as components in the GStreamer pipeline. The sensor module is used to obtain video stream and posture information in real time and output it to the target detection logic; The target detection logic is used to perform prediction and output positioning results; The embedded device includes a CPU, a GPU and a high-speed cache controller, and the high-speed cache controller is used to execute a multi-threaded scheduling mode.

Images