In the fast-paced realm of artificial intelligence, Edge AI emerges as a transformative innovation, providing solutions to critical concerns such as security and cost-effectiveness. Yet, the implementation of Edge AI poses challenges, particularly in the context of edge devices lacking a robust software stack and ecosystem.
This guide aims to provide a free people detection application on RZBoard V2L to easily test and experience the powerful RZ/V2L microprocessor (MPU) with more accurate and fast performance achieved by AI optimization technologies.
Challenges with edge AI developmentEdge AI development needs a lot of trial and error to meet the performance requirements within the given hardware limitations. Moreover, using specialized processors such as NPU, MPU, etc. is more challenging because they have powerful performance but have limitations to support all models.
- Trade off between accuracy and computing resources: There is no perfect model to have 100% accuracy with super fast speed and low memory usage. You have to test a lot of models to check their performance on the target HW for sorting them as a starting candidate. Even after that you face numerous optimization options and problems with models.
- Gap between SOTA Models and the HW: State-of-the-art models usually tries to solve the lack of the previous models to help you reduce trial and errors, but with new and creative technologies. For example, a leading AI framework PyTorch currently incorporates approximately 2, 200 operators—a number steadily increasing over time. However, what we have physical HWs right now were designed years ago so sometimes it is not even possible to use the recent models because of the compatibility issues.
Nota AI is one of the key players to address edge AI development challenges by AI model optimization technologies and has made an AI model optimization platform NetsPresso for AI engineers. Nota AI made an internal tool to analyze the model with HWs to check the compatibility and modify the model itself to make it compatible and more efficient. NetsPresso already supports RZ/V2L with its optimized AI models.
Optimized People Detection Demo OverviewThe AI model used in this demo was developed with NetsPresso. As shown in the below video, the model with Nota AI’s NetsPresso shows much faster and accurate performance.
Try this simple people detection demo to start development with RZ/V2L and experience the real performance of MPUs.
1. Download SD Card ImageTry this simple people detection demo to start development with RZ/V2L and experience the real performance of MPUs.
2. How to Flash a file to SD CARDDownload link of a program that can flash SD cards: Balena Etcher
- 1) Open the Balena Etcher and select the downloaded file you want to write on the SD card (downloaded file: located in the sd_image folder).
- 2) Select the drive you want to write your image to (32GB of SD card).
- 3) Review your selections and click 'Flash!' to begin writing data to the SD card.
- 4) After the Flash is over, you are ready to use the SD card.
- 0) Write the bootloaders to the board ( Attached is the boot loader file used on the demo board). Check AVNet GitHub: How to write the boot loader to the board.
- 1) Insert the microSD card to the Board.
- 2) Change the SW1 setting as shown in the figure.
- 3) Connect the USB camera to the Board.
- 4) Connect the HDMI monitor to the Board.
- 5) Connect the power cable to the Board.
- 6) Press the power button for 1 second to turn on the board.
- 7) After the boot-up, the following screen will be displayed on the HDMI monitor.
Notes for Running the Demo Application: The board is encountering freezing issues every 3 hours due to a licensing problem. To resolve this freezing problem, the board has been configured to automatically reboot every two hours. Following each reboot, the demo application will initiate automatically. Here are the SSH access details for the device: ID: root / Password: 0000
ConclusionIn conclusion, the development of edge AI poses various challenges such as optimizing models and addressing compatibility issues. However, solutions like NetsPresso by Nota AI can help simplify the process. The demonstration of people detection showcases the capabilities of the RZ/V2L MPU, although occasional freezing may occur due to licensing problems. These issues are mitigated through automatic reboots. Despite these challenges, this guide empowers developers to harness the potential of edge AI by utilizing the efficient RZ/V2L microprocessor for effective solutions. With continuous advancements, edge AI holds the promise of enabling smarter and more efficient edge devices and applications, thereby fostering innovation in the field.
Connect with usHave questions or are ready to embark on your Edge AI journey? Contact us at contact@nota.ai. Let’s collaborate to optimize your applications and shape the future of AI together.
Comments