NessajCN / rkod

Utility of RKNN C API of Rockchip RK3588 written in Rust.

It's the Utility of Rockchip's RKNN C API on rk3588. Written in Rust with FFI. src/bindings.rs was generated by bindgen wrapper.h -o src/bindings.rs

This repo is actually a Rust port of the yolov8 example in rknn_model_zoo

To run it:

• Download yolov8n.onnx

git clone https://github.com/NessajCN/rkod.git
cd rkod/model
bash ./download_model.sh

• Convert onnx to rknn model following this instruction. Install any dependencies necessary.

cd python
python convert.py <onnx_model> <TARGET_PLATFORM> <dtype(optional)> <output_rknn_path(optional)>

# such as: 
python convert.py ../model/yolov8n.onnx rk3588

• Move yolov8.rknn in rkod/model

• Fill model/label_list.txt with object name labels as you trained(one per line). Example could be found in model/coco_80_labels_list.txt.

• Change the const OBJ_CLASS_NUM in src/od.rs to the total number of entries you put in model/labels_list.txt. e.g. const OBJ_CLASS_NUM: i32 = 80; if you adopted model/coco_80_labels_list.txt as the labels_list.

• To detect objects in a single image:

cargo run -- -m <model/path> -i <image/path>

• To continuously detect the rtsp stream:

cargo run -- -m <model/path> -i rtsp://<rtsp-stream-path>

Note: It is recommended to compile and install the Rockchip version of ffmpeg following this guild if you intend to detect an rtsp streaming.