FastDonkey : a c/c++ framework for Donkeycar

FastDonkey is a c/c++ framework for DonkeyCar to autonumously drive.

What's better to the python version donkeycar

• Higher loop frequency. Main loop cycle time is shortened from 30ms to 20ms roughly.
• PWM direct drive from Pi. Do not need a servo drive board. Also no I2C communication delay.

Dependency

• Python version donkeycar. Used for training data collection and keras model training.
• ncnn deep learning framework. For high performance inference computing. Install it on Pi.
• Raspicam. C++ API for using Raspberry camera. Install it on Pi.
• Keras2caffe. Transform model from keras to caffe. Install it on PC.
• RPIO. a GPIO toolbox for the Raspberry Pi. No need to install, source code included in this repo.
• caffe. Caffe is a deep learning framework. Install it on PC.

How to use

Prepare model.

1. Get keras model from python version Donkeycar.
2. Transform to caffe model with Keras2caffe in PC. Remember to change your model path in keras_2_caffe.py. This script is from this repo.

python keras_2_caffe.py

3. Upgrade caffe model in PC for ncnn. Upgrade tools are from caffe.

./upgrade_net_proto_binary.bin caffe_mypilot.caffemodel new_caffe_mypilot.caffemodel
./upgrade_net_proto_text.bin caffe_mypilot.prototxt new_caffe_mypilot.prototxt

4. Transorm updated caffe model to ncnn model in Pi. caffe2ncnn is from ncnn.

./caffe2ncnn new_caffe_mypilot.prototxt new_caffe_mypilot.caffemodel ncnn_mypilot.param ncnn_mypilot.bin

Wiring

1. Just connect any pwm pins to throttle and steering servo. My connection is as follows:

throttle， 18(RGPIO number）,12(pin number)，G1(gpio number)
steering， 12(RGPIO number）,32(pin number)，G26(gpio number)

Compile and RUN

1. clone this repo and :

mkdir build
cd build
cmake ..
make
sudo ./fastdonkey

Enjoy crashing! :-P

Licence What ever licence

TODOs:

1. Put getting frame from camera to another thread. May save another 10ms.
2. Use ncnn optimize tool to accelerate inference speed.

---

广告时间: 如果你没有3D打印机，可以考虑在这个淘宝店购买3D打印支架