Simultaneous Control of 4 YOLOv8 Video Streams via MYIR's RK3576 Board

In the era of rapid technological advancement, the integration of artificial intelligence and edge computing is transforming our lives at an unprecedented pace. The RK3576 processor features a quad-core Cortex-A72 and a quad-core Cortex-A53 for foundational computing power, coupled with a 6TOPS NPU for model inference tasks. Utilizing the YOLOv8 model is also straightforward, so let's examine its capabilities using MYIR's MYD-LR3576 RK3576 Development Board.

About YOLO

YOLO (You Only Look Once) is a leading real-time object detection algorithm series in the industry, known for its perfect balance of speed and accuracy. Since its release, it has undergone several version changes, and the following figure illustrates its development history.

Evolution of YOLO Versions

YOLOv8 achieves an optimal balance by seamlessly integrating high performance, user-friendly design, cutting-edge architectural modernity, and a robust ecosystem. This makes it the most comprehensive and hassle-free choice available in the current market, offering unparalleled ease and efficiency for users seeking advanced object detection capabilities.  

YOLOv8 has numerous suffixes, which are simply listed in a table to indicate their respective meanings:

Table 1-1 By task type

Table 1-2 By model size

MYIR Demo Model Selection

For MYD-LR3576, the s/n small model with fewer parameters is more suitable. We demonstrate this by using basic functions and the -seg, -obb, and -pos parameters.

Test video scenes separately. The implementation is as follows:

The MYD-LR3576 features three MIPI-CSI interfaces, with three MY-CAM004M MIPI Camera modules connecting to each channel. It supports four AHD HD cameras using a 2+1+1 configuration, outputting H.264-encoded RTSP streams at 1920×1080 resolution and 30 fps. After processing by the MYD-LR3576 development board, each channel delivers 1920×1080 video at 25 fps, while the combined output of four channels reaches 60-70 fps. The CPU utilization approaches 100%, with the NPU achieving a 50-60% utilization rate.

Connections

Camera data processing flowchart

What optimizations does MYIR implement in the program?

AI inference round

• Get CSI frame data

• Crop data to a smaller image of xxx*yyy

• Process using RKNN API

• Get feature location and similarity

• Enlarge to original image

• Add boxes and similarity values to the original image

This results in low CPU utilization, reduced video capture frame rates, and ultimately a lagging display.

MYIR utilizes a thread pool architecture to process the aforementioned operations via threads, fully exploiting the capabilities of 4x A72 and 4x A53 cores while employing the RGA for image cropping and scaling. The system harmonizes the CPU, GPU, NPU, and VPU modules to optimize resource utilization.

The RK3576 demonstrates exceptional performance in the YOLOv8 model, with applications spanning multiple domains. For instance, in smart security systems deployed at public venues like airports, train stations, and shopping malls, it swiftly identifies abnormal behaviors such as fights or fleeing individuals, triggering instant alerts. Through facial recognition technology, the system verifies identities of visitors against its database, enabling real-time monitoring and tracking of key personnel. Additionally, it can provide rapid feedback when integrated with intelligent robots.

Stay tuned for more innovative applications of MYD-LR3576.