CN117830857A - A space target detection method and related equipment based on Atlas200DK - Google Patents

Abstract

The invention discloses a space target detection method based on Atlas200DK and related equipment, the method is based on Atlas200DK, resnet18 is adopted as a backbone network to replace an original feature extraction network Darknet53 in a YOLO v3 network model, the combination greatly improves the detection rate of the model, meanwhile, the problem of suitability between an algorithm and a processor is considered, the target detection model is subjected to model file conversion processing to ensure that the optimized algorithm can be normally executed, finally, an obtained space target image is input into the target detection model supported by Atlas200DK, a target detection result is obtained through analysis processing, and the reasoning speed of the model can be ensured through the optimization and the suitability adjustment of the YOLO v3 network model even when facing a large amount of data, so that the processing effect of high speed and low power consumption is achieved; the method can improve the processing speed and the detection precision of the space target monitoring, and meets the requirement of the current spaceborne computer system on the high reliability of the space target monitoring.

Description

A space target detection method and related equipment based on Atlas200DK

Technical Field

The invention belongs to the field of satellite-borne computers, and in particular relates to a space target detection method based on Atlas200DK and related equipment.

Background technique

Space target detection is an important part of the field of space target monitoring. Different target detection algorithms are suitable for different application scenarios. Among them, the YOLO series of target detection algorithms have been widely used due to their efficient detection performance.

With the development of the times, satellite-borne computer systems have put forward higher requirements for the reliability of space target monitoring. Generally speaking, there is a lot of space image information and the amount of calculation in the process of space target identification is also large.

Currently, the use of computing devices such as CPU and GPU, due to the use of traditional YOLO series algorithms, has problems such as slow processing speed and high power consumption in data processing, which obviously cannot meet the high reliability requirements of satellite-borne computer systems for monitoring space targets.

Summary of the invention

In order to overcome the shortcomings of the above-mentioned technology, the present invention provides a space target detection method and related equipment based on Atlas200DK, which can solve the technical problems of slow processing speed and high power consumption of existing computing devices such as CPU and GPU when the data processing volume is large.

In order to achieve the above object, the present invention adopts the following technical solution:

A space target detection method based on Atlas200DK, applied to Atlas200DK, includes:

Acquire space target images on Atlas200DK;

Input the space target image into the target detection model supported by Atlas200DK, and obtain the target detection result after parsing and processing;

Among them, the target detection model supported by Atlas200DK is obtained by converting the target detection model through model file conversion; the target detection model is a model structure constructed by YOLO v3 and Resnet18, which is obtained after training and testing.

Furthermore, the step of constructing the target detection model includes:

Replace the feature extraction network in the YOLO v3 network model with the Resnet18 network to obtain the YOLO-R network structure.

Furthermore, the specific steps of converting the model file of the target detection model include:

Convert the target detection model into an om model file and output the target detection model supported by Atlas200DK.

Furthermore, the target detection model adopts the prototxt model file under the Caffe framework.

Furthermore, during the model file conversion process, the prototxt model file under the Caffe framework is customized for network modification and AIPP configuration.

Furthermore, the auxiliary development tool ATC in the Ascend AI processor software stack is used to convert the target detection model into an om model file, wherein the model conversion is completed by executing the ATC command, and Mind Studio is used to proofread the conversion result.

Furthermore, the acquired spatial target image is input into the target detection model supported by Atlas200DK to obtain the inference detection result, which is then analyzed and processed by bounding box screening in combination with IoU and non-maximum suppression to obtain the target detection result.

A space target detection system based on Atlas200DK, used to implement the steps of the above-mentioned space target detection method based on Atlas200DK, comprising:

Image acquisition module, used to acquire space target images on Atlas200DK;

The target detection module is used to input the acquired space target image into the target detection model supported by Atlas200DK, and obtain the target detection result through parsing and processing;

Among them, the target detection model supported by Atlas200DK is obtained by converting the target detection model through model file conversion; the target detection model is a model structure constructed by YOLO v3 and Resnet18, which is obtained after training and testing.

A device comprising:

Memory for storing computer programs;

A processor is used to implement the steps of the above-mentioned space target detection method based on Atlas200DK when executing the computer program.

A computer-readable storage medium stores a computer program, which is used to implement the steps of the above-mentioned space target detection method based on Atlas200DK when executed by a processor.

Compared with the prior art, the present invention has the following beneficial effects:

The present invention also provides a space target detection method based on Atlas200DK. The method is based on Atlas200DK, uses Resnet18 as a backbone network, and replaces the original feature extraction network Darknet53 in the YOLO v3 network model. This combination greatly improves the detection rate of the model. At the same time, the compatibility problem between the algorithm and the processor is considered. By converting the target detection model into a model file, it is ensured that the optimized algorithm can be executed normally. Finally, the acquired space target image is input into the target detection model supported by Atlas200DK, and the target detection result is obtained through parsing. By optimizing the YOLO v3 network model and adjusting the adaptability, the reasoning speed of the model can still be guaranteed even in the face of a large amount of data, and the processing effect of high speed and low power consumption is achieved. The method can improve the processing speed and detection accuracy of space target monitoring, and meet the high reliability requirements of the current onboard computer system for space target monitoring.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a framework diagram of a space target detection method flow based on Atlas200DK provided by an embodiment of the present invention;

FIG2 is a diagram of a YOLO-R network structure provided by an embodiment of the present invention;

FIG3 is a flow chart of a space target detection method based on Atlas200DK provided by the present invention;

FIG. 4 is a schematic diagram of the structure of a space target detection system based on Atlas200DK provided by the present invention.

Detailed ways

The present invention provides a space target detection method based on Atlas200DK, as shown in FIG3 , comprising the following steps:

S1: Acquire space target images on Atlas200DK.

S2: Input the space target image into the target detection model supported by Atlas200DK, and obtain the target detection result after parsing.

Among them, the target detection model supported by Atlas200DK is obtained by converting the target detection model through model file conversion; the target detection model is a model structure constructed by YOLO v3 and Resnet18, which is obtained after training and testing.

The steps for building the above target detection model include:

Replace the feature extraction network in the YOLO v3 network model with the Resnet18 network to obtain the YOLO-R network structure.

The specific steps of converting the model file of the target detection model include:

Convert the target detection model into an om model file and output the target detection model supported by Atlas200DK.

Here, the target detection model uses the prototxt model file under the Caffe framework.

During the model file conversion process, the prototxt model file under the Caffe framework is customized for network modification and AIPP configuration.

Specifically, the auxiliary development tool ATC in the Ascend AI processor software stack is used to convert the target detection model into an om model file. The model conversion is completed by executing the ATC command, and Mind Studio is used to proofread the conversion result.

Specifically, the acquired spatial target image is input into the target detection model supported by Atlas200DK to obtain the inference detection result, the inference detection result is analyzed and bounding box screening is performed in combination with IoU and non-maximum suppression to obtain the target detection result.

As shown in FIG4 , the present invention further provides a space target detection system based on Atlas200DK, including: an image acquisition module, used to acquire a space target image on Atlas200DK; a target detection module, used to input the acquired space target image into a target detection model supported by Atlas200DK, and obtain a target detection result through parsing; wherein, the target detection model supported by Atlas200DK is obtained by converting the target detection model through a model file; and the target detection model is obtained through a model structure constructed by YOLO v3 and Resnet18 after training and testing.

The present invention also provides a device, comprising: a memory for storing a computer program; and a processor for implementing the steps of the space target detection method based on Atlas200DK when executing the computer program.

When the processor executes the computer program, the above-mentioned steps of space target detection based on Atlas200DK are implemented, for example: obtaining a space target image on Atlas200DK; inputting the obtained space target image into a target detection model supported by Atlas200DK, and obtaining a target detection result through parsing; wherein, the target detection model supported by Atlas200DK is obtained by converting the target detection model through a model file; and the target detection model is obtained through a model structure constructed by YOLO v3 and Resnet18 after training and testing.

Alternatively, when the processor executes the computer program, the functions of each module in the above system are realized, for example: an image acquisition module, used to acquire a space target image on Atlas200DK; a target detection module, used to input the acquired space target image into a target detection model supported by Atlas200DK, and obtain a target detection result through parsing; wherein, the target detection model supported by Atlas200DK is obtained by converting the target detection model through a model file; the target detection model is a model structure constructed by YOLO v3 and Resnet18, and is obtained after training and testing.

Exemplarily, the computer program may be divided into one or more modules/units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules/units may be a series of computer program instruction segments capable of completing preset functions, and the instruction segments are used to describe the execution process of the computer program in the space target detection device based on Atlas200DK. For example, the computer program may be divided into an image acquisition module and a target detection module; the specific functions of each module are as follows:

The image acquisition module is used to acquire the space target image on Atlas200DK; the target detection module is used to input the acquired space target image into the target detection model supported by Atlas200DK, and obtain the target detection result after parsing. Among them, the target detection model supported by Atlas200DK is obtained by converting the target detection model through the model file; the target detection model is a model structure constructed by YOLO v3 and Resnet18, which is obtained after training and testing.

The space target detection device based on Atlas200DK can be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The space target detection device based on Atlas200DK may include, but is not limited to, a processor and a memory. Those skilled in the art will appreciate that the above is an example of a space target detection device based on Atlas200DK, and does not constitute a limitation on the space target detection device based on Atlas200DK, and may include more components than the above, or a combination of certain components, or different components. For example, the space target detection device based on Atlas200DK may also include input and output devices, network access devices, buses, etc.

The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The processor is the control center of the space target detection based on Atlas200DK, and uses various interfaces and lines to connect various parts of the entire space target detection device based on Atlas200DK.

The memory can be used to store the computer program and/or module, and the processor implements various functions of the space target detection device based on Atlas200DK by running or executing the computer program and/or module stored in the memory, and calling the data stored in the memory.

The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the space target detection method based on Atlas200DK are implemented.

If the module/unit integrated in the space target detection system based on Atlas200DK is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

Based on such understanding, the present invention implements all or part of the processes in the above-mentioned space target detection method based on Atlas200DK, and can also be completed by instructing related hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and when the computer program is executed by a processor, the steps of the above-mentioned space target detection method based on Atlas200DK can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or preset intermediate form, etc.

The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

It should be noted that the content contained in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable storage media do not include electrical carrier signals and telecommunication signals.

The present invention will be further described below in conjunction with the embodiments and drawings:

Example

As mentioned in the background technology: In the existing satellite computer systems, computing devices such as CPU and GPU are used. Due to the use of traditional YOLO series algorithms, there are problems such as slow processing speed and high power consumption in data processing. Obviously, it cannot meet the high reliability requirements of the satellite computer system for space target monitoring.

In order to solve the above problems, this embodiment provides a space target detection method based on Atlas200DK. This method mainly uses an improved YOLO v3 algorithm for target detection. The improved YOLO v3 algorithm has efficient detection performance. Through Atlas 200DK, the reasoning speed of the algorithm can be further improved to complete efficient detection tasks.

The main steps of this detection method are: divided into a model training phase and a model reasoning phase; as shown in Figure 1, the model training phase is mainly implemented on the server side to obtain a target detection model. The present invention is mainly optimized for the Ascend AI processor, and the feature extraction network of YOLO v3 is replaced with Resnet18 to complete the optimization. The model reasoning phase is implemented on Huawei Atlas200DK. Based on this, the main idea of the present invention is:

(1) Algorithm Optimization

Atlas 200DK is a hardware product in the form of a developer kit. It uses the Ascend 310 chip as the core of the acceleration module. The Ascend 310 chip improves the algorithm performance through its powerful computing power. Therefore, when designing the algorithm, we consider how to make full use of the computing unit of the new Ascend AI series chips. In this algorithm optimization process, the algorithm optimization is mainly achieved by modifying the network structure.

The target detection algorithm based on YOLO v3 used in this embodiment has a feature extraction network of Darknet53. Due to the large amount of calculation data, the traditional YOLO v3 calculation speed is relatively slow and cannot meet the high reliability requirements of the current onboard computer system for space target monitoring. Therefore, in order to improve the detection rate of the model, in this embodiment, the backbone network Darknet53 of YOLO v3 is replaced with a shallower backbone network to complete feature extraction. After completing the study of various cases of using Atlas200DK for application development in the Ascend community, a typical classic case is to use Resnet18 to complete target recognition and detection tasks, and it can be seen from various cases that Resnet18 has good performance, which is one of the invention points of the present invention. In the actual application implementation process of Atlas200DK, the first thing to consider is the compatibility of the algorithm with the Ascend AI processor. If there is a situation where it cannot be adapted, it will bring great resistance to the subsequent application implementation. At the same time, Darknet53 network is an important variant network of Resnet. In order to ensure that the YOLOv3 target detection algorithm can be executed normally after optimization, it is considered to gradually replace Darknet53 with a Resnet series network structure similar to its structure, namely Resnet18.

As shown in Figure 2, the above optimization method is used to optimize YOLO v3 to obtain the structure of YOLO-R. When using this structure for target detection of the algorithm, its process is similar to the original YOLO v3. First, it is necessary to use the Resnet18 network to complete feature extraction, extract features of three different scales respectively, and then use convolutional networks and upsampling to complete the output of target results of different scales. Next, the target box is predicted and useful boxes are extracted from the features. The target box prediction here is the same as the original YOLO v3 network. Finally, the final result box is screened out according to the non-maximum suppression algorithm.

(2) Overall configuration of the experimental platform

In the model inference stage in Atlas200DK, the original model file needs to be converted into an offline network model om model file supported by the Ascend AI processor. In the Ascend AI processor, model files under the current mainstream deep learning frameworks are supported, such as Caffe, MindSpore, TensorFlow, etc. This embodiment uses the model file prototxt file and the weight file caffemodel under the Caffe framework to successfully convert the model through the model conversion tool.

After completing the model conversion in the model reasoning stage, application development is performed according to the development process of Atlas200DK and in combination with the ACL development API. The general application development process is: preprocessing of media data, loading the model for model reasoning, parsing the model reasoning results, and finally processing the obtained reasoning results. After the application development is completed, the model can be deployed and run according to the development language of different applications. In the current version of the Atlas200DK developer kit, APIs are provided in C++ version and python version, so the application development language is mainly C++ or python. This embodiment uses C++ language for model deployment and model operation. The compilation of the project files is completed in the development environment first, and then deployed and run on the developer board.

(3) Algorithm and application development and implementation

In Atlas200DK, the existing model must be converted into an offline model om file supported by the Ascend AI processor before it can be used for subsequent application development and business implementation and completion. When performing model conversion, the auxiliary development tool ATC in the Ascend AI processor software stack is mainly used. The tool provides users with two different model conversion methods, namely the model conversion tool of MindStudio and the ATC command line for model conversion. When performing model conversion, customized network modifications and AIPP configurations need to be performed as needed. Customized network modifications are mainly performed on the model file (*.prototxt) under the Caffe framework to meet the operator support during model conversion of the Ascend AI processor. Both model conversion operations require the modification of the file content before model conversion, and the AIPP configuration provides different configuration methods for two different conversion methods. This embodiment uses ATC as an auxiliary development tool for model conversion. When performing model conversion through the ATC command line, put the model file, weight file, and AIPP configuration file in the same folder, open the terminal in the folder for environment configuration, and complete the model conversion through the ATC command line. After using ATC to convert the model, use MindStudio to check whether the structure of the om model is consistent with the original caffe model. The above is another inventive point of the present invention.

According to the analysis of various functions in the application scenario analysis, a sub-function is created for each function to implement the function of each scenario.

1) Resource initialization and deinitialization: When using ACL for program development, various resources must be initialized. ACL initialization is implemented through the acl.init interface, and the interface under acl.media is used to complete the data preprocessing resource application. The interface model under acl.mdl is used to load the resource application to avoid problems in subsequent development operations. When a resource is used, it needs to be deinitialized to release the resource.

2) Model inference function: mainly calls the interface under acl.mdl. After completing the preprocessing operation of the image, the image can be sent to the model for model loading and execution to complete the model inference task. After the model inference is completed, the model resources need to be released.

3) Data post-processing function: After the model inference operation is completed, the image inference results are obtained and the model inference results are parsed to obtain the final target detection results. Different original models will make the post-processing functions different. For example, in this article, the YOLOv3 model encapsulates parameters such as confidence and non-maximum suppression during model conversion. Therefore, there is no need to process this part in post-processing. You only need to obtain the inference results and output the results. The YOLO-R model requires a detailed analysis of the inference results. It is necessary to parse the model inference results in the post-processing function, and combine IoU (intersection over union measurement) and non-maximum suppression to complete bounding box screening related operations to obtain the final target detection results. After obtaining the inference results, output the inference detection results, and use opencv or pillow functions on the original image to mark the inference results.

In summary, the present invention provides a space target detection method based on Atlas200DK. The method is based on Atlas200DK, uses Resnet18 as the backbone network, and replaces the original feature extraction network Darknet53 in the YOLO v3 network model. This combination greatly improves the detection rate of the model. At the same time, the compatibility problem between the algorithm and the processor is considered. The trained target detection model is converted into a model file to ensure that the optimized algorithm can be executed normally. Finally, the acquired space target image is input into the target detection model supported by Atlas200DK, and the target detection result is obtained through parsing. By optimizing the YOLO v3 network model and adjusting its adaptability, the reasoning speed of the model can still be guaranteed even in the face of a large amount of data, achieving a high-speed and low-power processing effect. The method can improve the processing speed and detection accuracy of space target monitoring, and meet the high reliability requirements of the current onboard computer system for space target monitoring.

The above embodiment is only one of the implementation methods that can realize the technical solution of the present invention. The scope of protection claimed by the present invention is not limited only to this embodiment, but also includes changes, replacements and other implementation methods that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention.

Claims (10)

1. The spatial target detection method based on Atlas200DK is characterized by being applied to Atlas200DK and comprising the following steps:

acquiring a space target image on Atlas200 DK;

inputting a space target image into a target detection model supported by Atlas200DK, and obtaining a target detection result through analysis;

the target detection model supported by Atlas200DK is obtained by converting a model file from the target detection model; the target detection model is obtained by training and testing a model structure constructed by YOLO v3 and Resnet18.

2. The Atlas200 DK-based spatial target detection method according to claim 1, wherein the step of constructing the target detection model includes:

and replacing the feature extraction network in the YOLO v3 network model with a Resnet18 network to obtain a YOLO-R network structure.

3. The Atlas200 DK-based spatial target detection method according to claim 1, wherein the specific step of performing model file conversion processing on the target detection model comprises:

and converting the target detection model into an om model file, and outputting the target detection model supported by Atlas200 DK.

4. The spatial target detection method based on Atlas200DK according to claim 3, wherein the target detection model adopts a prototxt model file under a Caffe framework.

5. The Atlas200 DK-based spatial target detection method of claim 4, wherein the custom network modification and AIPP configuration are performed on the prototxt model file under the Caffe framework at the same time during the model file conversion process.

6. The Atlas200 DK-based spatial object detection method of claim 3, wherein the object detection model is converted into an om model file by using an auxiliary development tool ATC in a software stack of a lifting AI processor, wherein the model conversion is completed by executing an ATC command, and the conversion result is collated by using a mini Studio.

7. The Atlas200 DK-based spatial target detection method of claim 1, wherein the acquired spatial target image is input to a target detection model supported by Atlas200DK to obtain an inference detection result, and the inference detection result is analyzed and combined with IoU and non-maximum suppression to perform bounding box screening processing to obtain the target detection result.

8. Atlas200 DK-based spatial target detection system for implementing the steps of the Atlas200 DK-based spatial target detection method of any one of claims 1-7, comprising:

the image acquisition module is used for acquiring a space target image on Atlas200 DK;

the target detection module is used for inputting the acquired space target image into a target detection model supported by Atlas200DK, and obtaining a target detection result through analysis;

the target detection model supported by Atlas200DK is obtained by converting a model file from the target detection model; the target detection model is obtained by training and testing a model structure constructed by YOLO v3 and Resnet18.

9. An apparatus, comprising:

a memory for storing a computer program;

a processor for implementing the Atlas200 DK-based spatial target detection method of any one of claims 1-7 when executing said computer program.

10. A computer readable storage medium storing a computer program, characterized in that the computer program is executed by a processor for implementing the steps of the Atlas200 DK-based spatial target detection method of any of claims 1-7.