zyayoung / k210-tracking

Run yolov3 on Kendryte K210 (an AI capable SoC)

Kendryte K210 Animal Tracking

Kendryte Standalone SDK

Demo

Analysis tool

Now available TrackingAnalyze

Train

Train yolo model using darknet or this repo.

To conver your keras model to kmodel, MaixPy_scripts or nncase can be used.

First prepare your yolo.h5 file, then

tflite_convert --keras_model_file=yolo.h5 --output_file=yolo.tflite
./tflite2kmodel.sh yolo.tflite

Usage

Compile

• Ubuntu

Download RISC-V 64bit toolchain for Kendryte K210_ubuntu_amd64 from https://kendryte.com/downloads/.

Extract it to /opt/riscv-toolchain, and add export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/riscv-toolchain/bin/ to bashrc.

mkdir build && cd build
cmake .. -DPROJ=yolo3_frame_test_public_maixpy -DTOOLCHAIN=/opt/riscv-toolchain/bin && make

• Windows

Download and install CMake and the latest toolchain.

mkdir build && cd build
cmake -G "MinGW Makefiles" .. -DPROJ=yolo3_frame_test_public_maixpy -DTOOLCHAIN=/path/to/toolchain/bin
make

flash

You will get yolo3_frame_test_public_maixpy.bin.

If you want to flash it in UOG, using yolo3_frame_test_public_maixpy.bin, then using flash-tool(s) burn it to your flash.

sudo kflash yolo3_frame_test_public_maixpy.bin -B dockE -p /dev/ttyUSB0 -b 3000000 -t

You may want to flash your bin and model toghther with kfpkg

cp ../src/yolo3_frame_test_public_maixpy/kfpkg/kpu_yolov3.kfpkg .
zip kpu_yolov3.kfpkg yolo.kmodel yolo3_frame_test_public_maixpy.bin
sudo kflash kpu_yolov3.kfpkg -B dockE -p /dev/ttyUSB0 -b 3000000 -t

Credit

• Yolov3
• kendryte-standalone-sdk
• keras-yolo3
• K210_Yolo_framework
• M. Lorbach, E. I. Kyriakou, R. Poppe, E. A. van Dam, L. P. J. J. Noldus, and R. C. Veltkamp, “Learning to Recognize Rat Social Behavior: Novel Dataset and Cross-Dataset Application,” Journal of Neuroscience Methods, 2017. data
• MaixPy_scripts

k210-tracking's Issues

这个网络能多跑多少FPS啊

请问这个网络在k210上能跑什么样的速度啊？

Pretrained model (or documenting how to train it)

Missing feature

The pretrained model, or documentation about it.

Justification

It'd be useful to have the pretrained model detecting people (as in the image in the readme).

• Partly so this code can be run quickly on the K210 to test e.g. FPS.
• Even better - so it can be used directly for detecting people (or whatever else the model detects).

Workarounds

Alternatively, documentation of how the model was trained. The current docs imply it's yolo, but here implies it's 'modilenet yolov3 alpha=0.75' which I think may come from https://github.com/zhen8838/K210_Yolo_framework ? Making this clearer (and the data used, etc.) would be an easier step than making the model available.