This project demonstrates YOLO (You Only Look Once) object detection using AXCL (Axera Computing Library) on Axera AX8850 devices with hardware acceleration.

Overview

The AXCL YOLO project provides AI-powered object detection capabilities using the YOLO algorithm, leveraging the NPU (Neural Processing Unit) of Axera AX8850 devices for high-performance inference.

Features

Hardware-Accelerated AI Inference: Utilizes Axera AX8850 NPU for fast YOLO object detection
Multiple YOLO Model Support: Compatible with various YOLO model formats
Real-time Processing: Optimized for real-time object detection applications
OpenCV Integration: Built-in image processing and visualization capabilities
Flexible Input Sources: Support for various input formats and sources

Prerequisites

Axera AX8850 device with NPU support
ARM64 cross-compilation toolchain (aarch64-none-linux-gnu-)
AXCL MSP (Media Signal Processing) libraries
OpenCV 4.5.5 libraries
YOLO model files (trained models in compatible format)

Building

Configure the build environment:

# The project uses SCons build system
# Ensure your toolchain is properly configured in config_defaults.mk

The build system will automatically clone the required axcl-samples repository:

# This happens automatically during build
git clone https://github.com/dianjixz/axcl-samples.git

Build the project:
```
scons
```

Architecture Support

ARM64 (aarch64): Full support with OpenCV 4.5.5 ARM64 libraries
x86_64: Development support with AVX2 optimization

Dependencies

Core Libraries

AXCL Runtime: Core AXCL runtime libraries (axcl_rt)
AXCL MSP: Media signal processing libraries (axcl_msp)
AXCL Communication: Device communication libraries (axcl_comm, axcl_pcie_msg, axcl_pcie_dma)
AXCL System: System management libraries (axcl_sys)

Third-Party Libraries

OpenCV 4.5.5: Image processing and computer vision
- opencv_core, opencv_highgui, opencv_imgcodecs, opencv_imgproc, opencv_videoio
Image Format Support: TIFF, OpenJPEG2, JPEG-Turbo, WebP, PNG
Logging: spdlog for structured logging
Threading: pthread for multi-threading support

Additional Dependencies

Intel TBB: Threading Building Blocks (libittnotify.a)
Compression: zlib for data compression
System Libraries: dl (dynamic linking), tegra_hal

Usage

The project builds multiple executables based on the source files in the axcl-samples repository. Each executable corresponds to a specific YOLO implementation or demo.

Basic Usage Pattern

# Run a YOLO detection application
./<yolo_app_name> [options]

Common Parameters

Model file path
Input image/video path
Output directory
Confidence threshold
NMS (Non-Maximum Suppression) threshold
Device selection

Model Support

The project supports various YOLO model formats optimized for Axera hardware:

YOLOv3, YOLOv4, YOLOv5, YOLOv8
Custom trained models in compatible formats
Models optimized for AX8850 NPU architecture

Performance Optimization

NPU Acceleration: Leverages dedicated neural processing unit
Memory Management: Optimized memory allocation for real-time processing
Pipeline Processing: Efficient data flow for continuous inference
Multi-threading: Parallel processing for improved throughput

Configuration

The project uses several configuration files:

config_defaults.mk: Build-time configuration
axcl.json: Runtime AXCL configuration (if needed)
Model-specific configuration files

Hardware Requirements

Target Platform: Axera AX8850
Architecture: ARM64 (aarch64)
NPU: Required for AI inference acceleration
Memory: Sufficient RAM for model loading and inference
Storage: Space for model files and temporary data

Development Notes

The project automatically downloads the axcl-samples repository during build
OpenCV libraries are architecture-specific (ARM64 vs x86_64)
FFmpeg integration available for video processing workflows
RPATH configured for proper library loading (/usr/lib/axcl/ffmpeg)

Important Notes

⚠️ OpenCV Video Library Limitation: The OpenCV library included with the AXCL MSP SDK does not support video processing functionalities. If you want to use video-related examples such as ax_yolo11_steps_video, you need to execute the following command in the build terminal:

export CONFIG_USE_HOST_OPENCV4LIB=y

This will enable the use of the host's OpenCV library with video support during compilation.

To enable video functionality, please follow these steps:

Set the environment variable CONFIG_USE_HOST_OPENCV4LIB=y
Install libopencv4-dev on your host machine
Ensure that the host OpenCV library is compatible with your target architecture

Troubleshooting

Build Issues:
- Ensure proper toolchain configuration
- Verify OpenCV library paths for your architecture
- Check network connectivity for repository cloning
Runtime Issues:
- Verify AXCL libraries are properly installed
- Ensure model files are in correct format and location
- Check device permissions and NPU availability
Performance Issues:
- Monitor NPU utilization
- Optimize input resolution for your use case
- Adjust batch size and threading parameters

Examples

The axcl-samples repository contains various example implementations:

Basic YOLO object detection
Real-time video processing
Batch image processing
Custom model integration

External Dependencies

axcl-samples: https://github.com/dianjixz/axcl-samples.git
OpenCV 4.5.5: Computer vision library
Various image format libraries (TIFF, JPEG, PNG, WebP)