Target detection system and method based on low-power AI chip

Abstract

The invention relates to a target detection system and method based on a low-power AI chip, the system comprises an AI processing chip, a power supply unit, an interface unit and an onboard storage unit, the AI processing chip is respectively connected with the power supply unit, the interface unit and the onboard storage unit, and the AI processing chip is used for information processing; the power supply unit is used for supplying power; the interface unit is used for providing various interfaces; the onboard storage unit is used for storing data. An AI intelligent target detection technology is fused with an AI processing chip, a low-power-consumption edge end reasoning acceleration software and hardware system with certain computing power is constructed, a low-power-consumption AI chip serves as a core device, external video stream data are collected through input interfaces such as a USB and a network, and detection of a specified target is completed in the AI chip through a target detection algorithm; the artificial intelligence technology can be explored and innovated in the specialized application field.

Description

Technical Field

[0001] The present invention belongs to the field of AI chip control, and in particular to a target detection system and method based on a low-power AI chip. Background Art

[0002] Target recognition refers to the process of distinguishing a specific target (or a type of target) from other targets (or other types of targets). It includes both the recognition of two very similar targets and the recognition of one type of target from other types of targets.

[0003] The basic principle of target recognition is to use the target feature information such as amplitude, phase, spectrum and polarization in the radar echo, and estimate the size, shape, weight and physical characteristic parameters of the surface layer of the target through various mathematical multi-dimensional space transformations. Finally, the identification decision is made in the classifier based on the discrimination function determined by a large number of training samples.

[0004] In an information environment, improving the ability to perceive the global situation can effectively improve the ability to control the overall situation. The identification and positioning of specific targets are key technologies that affect the global situation perception. Therefore, using artificial intelligence as a technical means and taking perspectives such as sea, land, and air to solve the key problems of multi-dimensional specific target identification is of great significance to the generation and analysis of the global situation. Therefore, it is highly practical and urgent to carry out edge target detection development.

[0005] At present, the engineering applications of edge target detection mostly use NVIDIA smart modules as the hardware platform for technical development. There are two significant problems. First, it cannot meet the current application needs in the domestic context, and development is easily hindered due to product embargoes. Second, NVIDIA sells finished AI modules to the outside world, and the power consumption is relatively high. It cannot meet the rigid index requirements of size, weight, power consumption, etc. of miniaturized platforms such as drones. The development of domestic AI chips is relatively lagging, and the software ecological environment is relatively weak, which is also a technical barrier that the present invention needs to break through from the software technology level. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a target detection system and method based on a low-power AI chip. The low-power AI chip is used as the core device, external video stream data is collected through input interfaces such as USB and network, and the detection of specified targets is completed through a target detection algorithm inside the AI ​​chip.

[0007] The present invention solves the technical problem by adopting the following technical solutions:

[0008] The target detection system based on a low-power AI chip includes an AI processing chip, a power supply unit, an interface unit and an onboard storage unit, wherein the AI ​​processing chip is respectively connected to the power supply unit, the interface unit and the onboard storage unit, wherein the AI ​​processing chip is used for information processing; the power supply unit is used for power supply; the interface unit is used for providing multiple interfaces; and the onboard storage unit is used for storing data.

[0009] Moreover, it also includes the peripheral circuit of the AI ​​processing chip, wherein the peripheral circuit of the AI ​​processing chip includes an RTC clock, a WD unit and WIFI, the AI ​​processing chip is connected to the RTC clock via I2C, the AI ​​processing chip is connected to the WD unit via GPIO, and the AI ​​processing chip is connected to WIFI via PCIE.

[0010] Moreover, the power supply unit includes a Gigabit network interface, a 232 interface, a USB interface, a 422 interface and an HDMIOUT interface.

[0011] Moreover, the AI ​​processing chip adopts the low-power AI chip RK3588 as the core processor; the core processor is expanded with an EMMC chip for transplanting the on-chip operating system and storing data. After the system is powered on, the BOOT program will read the operating system information from the EMMC chip to complete the startup of the on-chip operating system.

[0012] A detection method for a target detection system based on a low-power AI chip, comprising the following steps:

[0013] Step 1: The power supply unit is powered on, and the operating system on the AI ​​processing chip starts normally;

[0014] Step 2: read the local video stream or real-time video stream data from the onboard storage unit, and the user program starts to execute;

[0015] Step 3: The AI ​​processor reads the video stream data, extracts the video stream data frame, and performs detection through the constructed target detection software;

[0016] Step 4: The target detection results are displayed locally or transmitted externally through the target detection software inside the AI ​​processor; the data information generated by the user program is exchanged with the outside world through the serial port.

[0017] Moreover, the method for constructing the target detection software in step 3 includes the deployment of an operating system on an AI processing chip and the construction of a target detection system architecture.

[0018] Moreover, the specific implementation method of the operating system deployment on the AI ​​processing chip is: using Ubuntu20.04 as the on-chip operating system for porting, and burning the Ubuntu20.04 system into the EMMC onboard storage expanded outside the chip through the system burning system of the RK3588 chip. After the power supply unit is powered on, the RK3588 chip reads the operating system firmware in the EMMC to complete the porting and deployment of the operating system; at the same time, the installation and deployment of opencv and qt software are completed through online and offline methods, so that the AI ​​module has an application environment for subsequent secondary development by users.

[0019] Moreover, the target detection system architecture includes algorithm model conversion software, development support software and target detection system software, wherein the algorithm model conversion software includes AI model training software and algorithm core construction software; the development support software includes peripheral interface device driver software and target detection system support software, and the target detection system software includes video reading software and AI reasoning software.

[0020] Moreover, the peripheral interface device driver software includes RS422 driver, RS232 driver, Gigabit network driver, USB driver and HDMI driver;

[0021] The target detection system supporting software includes QT software, Open software and FFmpeg software;

[0022] The video reading software reads local video streams and network video streams;

[0023] AI reasoning software includes AI model training, AI model transformation and AI model reasoning.

[0024] The advantages and positive effects of the present invention are:

[0025] The present invention includes an AI processing chip, a power supply unit, an interface unit and an onboard storage unit, wherein the AI ​​processing chip is connected to the power supply unit, the interface unit and the onboard storage unit respectively, wherein the AI ​​processing chip is used for information processing; the power supply unit is used for power supply; the interface unit is used for providing multiple interfaces; and the onboard storage unit is used for storing data. The present invention integrates AI intelligent target detection technology with the AI ​​processing chip to construct a low-power edge-end inference acceleration hardware and software system with certain computing power, and uses a low-power AI chip as the core device, collects external video stream data through input interfaces such as USB and network, and completes the detection of designated targets through target detection algorithms inside the AI ​​chip; the present invention can realize the exploration and innovation of artificial intelligence technology in specialized application fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a system structure diagram of the present invention;

[0027] Figure 2 Flow chart of the method of the present invention

[0028] Figure 3 This is a diagram of the target detection software architecture of the present invention. DETAILED DESCRIPTION

[0029] The present invention is further described in detail below with reference to the accompanying drawings.

[0030] The construction idea of ​​the present invention is:

[0031] The AI ​​module detection system adopts the core architecture of AI processing chip + power supply unit + multiple input and output interfaces in hardware architecture, and uses multiple interfaces such as network port, USB, serial port, etc. to interact with the outside world. In terms of hardware architecture, it adopts the core architecture of AI processing chip + multi-level power supply unit + multiple input and output interfaces, and uses multiple interfaces such as network port, USB, serial port, etc. to interact with the outside world. In terms of power supply, a multi-level power supply strategy is adopted, which is suitable for harsh environments with a large fluctuation range of power supply voltage, and the size of the hardware board is only 110mm×87mm, which meets the application requirements of small edge devices such as small drones for AI intelligent computing platforms.

[0032] In terms of AI module detection methods, in order to solve the problem that multi-scale targets are difficult to detect in complex environments, a target detection algorithm based on a multi-scale detection frame is used for detection. This algorithm targets larger detection targets in the image. Through a large-scale detection frame, the target detection is completed in the deep semantic feature layer. For smaller targets to be detected, shallow target features and deep semantic features are first used for feature fusion, so that the target feature information and semantic feature information of small-sized targets are restored, and a small-scale target detection frame is used to complete the detection of small targets. Using a multi-scale detection frame to detect multi-scale targets in the image can effectively improve the accuracy of multi-scale target detection.

[0033] Object detection systems based on low-power AI chips, such as Figure 1 As shown, it includes an AI processing chip, a power supply unit, an interface unit and an onboard storage unit, wherein the AI ​​processing chip is connected to the power supply unit, the interface unit and the onboard storage unit respectively, wherein the AI ​​processing chip is used for information processing; the power supply unit is used for power supply; the interface unit is used for providing multiple interfaces; and the onboard storage unit is used for storing data.

[0034] The present invention also includes a peripheral circuit of an AI processing chip, wherein the peripheral circuit of the AI ​​processing chip includes an RTC clock, a WD unit and WIFI, the AI ​​processing chip is connected to the RTC clock via I2C, the AI ​​processing chip is connected to the WD unit via GPIO, and the AI ​​processing chip is connected to WIFI via PCIE.

[0035] The power supply unit includes a Gigabit network interface, a 232 interface, a USB interface, a 422 interface and an HDMIOUT interface.

[0036] The AI ​​processing chip uses the low-power AI chip RK3588 as the core processor; the core processor is expanded with an EMMC chip for porting the on-chip operating system and storing data. After the system is powered on, the BOOT program will read the operating system information from the EMMC chip to complete the startup of the on-chip operating system. After the operating system is started, the AI ​​algorithm is used to detect and identify targets on the read video stream data, depending on the type of video stream data, whether it is local video data or external camera data. The AI ​​chip's built-in NPU resources are used to accelerate the AI ​​algorithm, thereby improving the performance of target detection.

[0037] A detection method for a target detection system based on a low-power AI chip, such as Figure 2 As shown, the following steps are included:

[0038] Step 1: The power supply unit is powered on, and the operating system on the AI ​​processing chip starts normally.

[0039] Step 2: read the local video stream or real-time video stream data from the onboard storage unit, and the user program starts to execute.

[0040] Step 3: The AI ​​processor reads the video stream data, extracts the video stream data frames, and performs detection through the constructed target detection software.

[0041] The method for building target detection software includes the deployment of an operating system on an AI processing chip and the construction of a target detection system architecture.

[0042] The specific implementation method of the operating system deployment on the AI ​​processing chip is as follows: To complete the construction of the target detection system, it is essential to transplant the on-chip operating system and deploy target detection related software. Ubuntu20.04 is transplanted as the on-chip operating system. Through the system burning system of the RK3588 chip, the Ubuntu20.04 system is burned into the EMMC onboard storage of the chip. After the power supply unit is powered on, the RK3588 chip reads the operating system firmware in the EMMC to complete the transplantation and deployment of the operating system; at the same time, the installation and deployment of opencv and qt software are completed online and offline, so that the AI ​​module has an application environment for subsequent secondary development by users. At this point, the deployment of the on-chip basic software environment has been completed.

[0043] like Figure 3As shown in the figure, the target detection system architecture includes algorithm model conversion software, development support software and target detection system software. The algorithm model conversion software is deployed in a PC for training AI target detection algorithms; the development support software is used to provide an environment for the development of target detection system software; the target detection system software includes video reading software for reading external video stream data.

[0044] Among them, the algorithm model conversion software includes AI model training software and algorithm core construction software; the development support software includes peripheral interface device driver software and target detection system support software, and the target detection system software includes video reading software and AI reasoning software. The AI ​​reasoning software is responsible for converting the AI ​​model into the data type supported by RK3588 and calling NPU resources to complete the acceleration and forward reasoning process of the AI ​​algorithm model.

[0045] The peripheral interface device driver software includes RS422 driver, RS232 driver, Gigabit network driver, USB driver and HDMI driver.

[0046] The target detection system supporting software includes QT software, Open software and FFmpeg software.

[0047] In the target detection model transplantation part, the present invention builds a target detection model through the algorithm core construction software, and completes the training of the AI ​​algorithm model through the AI ​​model training software, so that it has the target detection capability. The algorithm model with AI intelligent detection capability is converted into a model data architecture that can be recognized by the RK3588 chip through the cross-compilation model conversion tool deployed on the PC side. In order to support different detection rates of target detection of the on-chip operating system, the model conversion tool can convert the algorithm model into data types such as INT8 and FP16. After the model conversion is completed, the model is transplanted to the AI ​​module hardware platform through the network data line to complete the transplantation of the target detection model.

[0048] For the development of on-chip target detection system software, the present invention reads the video data of an external USB or network camera through the opencv software, transfers the video stream data content, and then extracts the image frame in the video stream and sends it to the target detection system. In the target detection system, the NPU resource interface of RK3588 is called by the program to realize the accelerated processing capability of the AI ​​detection algorithm and improve the speed of target detection in the video frame. Through video overlay processing, the detection results are superimposed on the video stream, and the video results are displayed through the peripheral display device or pushed to the external device through rtsp push stream by the program. The development of the target detection system software is completed. In order to make full use of the NPU resources of RK3588, the utilization rate of the three NPU cores of RK3588 can be increased to 80% by using multi-threaded processing technology during program development, which effectively improves the detection rate of the target detection system.

[0049] The video reading software reads local video streams and network video streams;

[0050] AI reasoning software includes AI model training, AI model transformation and AI model reasoning.

[0051] Step 4: The target detection results are displayed locally or transmitted externally through the target detection software inside the AI ​​processor; the data information generated by the user program is exchanged with the outside world through the serial port.

[0052] The present invention includes an AI processing chip, a power supply unit, an interface unit and an onboard storage unit, wherein the AI ​​processing chip is connected to the power supply unit, the interface unit and the onboard storage unit respectively, wherein the AI ​​processing chip is used for information processing; the power supply unit is used for power supply; the interface unit is used for providing multiple interfaces; and the onboard storage unit is used for storing data. The present invention integrates AI intelligent target detection technology with the AI ​​processing chip to construct a low-power edge-end inference acceleration hardware and software system with certain computing power, and uses a low-power AI chip as the core device, collects external video stream data through input interfaces such as USB and network, and completes the detection of designated targets through target detection algorithms inside the AI ​​chip; the present invention can realize the exploration and innovation of artificial intelligence technology in specialized application fields.

[0053] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation manner. Any other implementation manners derived by those skilled in the art based on the technical solution of the present invention also fall within the scope of protection of the present invention.

Claims

1. The target detection system based on low-power AI chip is characterized by: It includes an AI processing chip, a power supply unit, an interface unit and an onboard storage unit, wherein the AI ​​processing chip is connected to the power supply unit, the interface unit and the onboard storage unit respectively, wherein the AI ​​processing chip is used for information processing; the power supply unit is used for power supply; the interface unit is used for providing multiple interfaces; and the onboard storage unit is used for storing data.

2. The target detection system based on a low-power AI chip according to claim 1, characterized in that: It also includes peripheral circuits of the AI ​​processing chip, wherein the peripheral circuits of the AI ​​processing chip include an RTC clock, a WD unit and WIFI, the AI ​​processing chip is connected to the RTC clock via I2C, the AI ​​processing chip is connected to the WD unit via GPIO, and the AI ​​processing chip is connected to WIFI via PCIE.

3. The target detection system based on a low-power AI chip according to claim 1, characterized in that: The power supply unit includes a Gigabit network interface, a 232 interface, a USB interface, a 422 interface and an HDMIOUT interface.

4. The target detection system based on a low-power AI chip according to claim 1, characterized in that: The AI ​​processing chip uses the low-power AI chip RK3588 as the core processor; the core processor is expanded with an EMMC chip for transplanting the on-chip operating system and storing data. After the system is powered on, the BOOT program will read the operating system information from the EMMC chip to complete the startup of the on-chip operating system.

5. A detection method for a target detection system based on a low-power AI chip according to any one of claims 1 to 4, characterized in that: The following steps are involved: Step 1: The power supply unit is powered on, and the operating system on the AI ​​processing chip starts normally; Step 2: read the local video stream or real-time video stream data from the onboard storage unit, and the user program starts to execute; Step 3: The AI ​​processor reads the video stream data, extracts the video stream data frame, and performs detection through the constructed target detection software; Step 4: The target detection results are displayed locally or transmitted externally through the target detection software inside the AI ​​processor; the data information generated by the user program is exchanged with the outside world through the serial port.

6. The detection method of the target detection system based on the low-power AI chip according to claim 5 is characterized in that: The method for constructing the target detection software in step 3 includes deploying an operating system on an AI processing chip and building a target detection system architecture.

7. The detection method of the target detection system based on the low-power AI chip according to claim 6 is characterized in that: The specific implementation method of the operating system deployment on the AI ​​processing chip is: use Ubuntu20.04 as the on-chip operating system for porting, and burn the Ubuntu20.04 system into the EMMC onboard storage expanded outside the chip through the system burning system of the RK3588 chip. After the power supply unit is powered on, the RK3588 chip reads the operating system firmware in the EMMC to complete the porting and deployment of the operating system; at the same time, the installation and deployment of opencv and qt software are completed through online and offline methods, so that the AI ​​module has an application environment for subsequent secondary development by users.

8. The detection method of the target detection system based on the low-power AI chip according to claim 6 is characterized in that: The target detection system architecture includes algorithm model conversion software, development support software and target detection system software, wherein the algorithm model conversion software includes AI model training software and algorithm core construction software; the development support software includes peripheral interface device driver software and target detection system support software, and the target detection system software includes video reading software and AI reasoning software.

9. The detection method of the target detection system based on the low-power AI chip according to claim 8, characterized in that: The peripheral interface device driver software includes RS422 driver, RS232 driver, Gigabit network driver, USB driver and HDMI driver; The target detection system supporting software includes QT software, Open software and FFmpeg software; The video reading software reads local video streams and network video streams; AI reasoning software includes AI model training, AI model transformation and AI model reasoning.

Images