Run the Car

What you will learn:

• How to run the autonomous car.

What you need to modify according to your own model:

1. Variables:

   • KERNEL_CONV is the net_name parameter in compile.sh, CONV_INPUT_NODE & CONV_OUTPUT_NODE are the input_nodes & output_nodes in quant.sh.

   • CUT_SIZE is the number of lines to cut in initial images(be the same in process_img.py). STORESIZE_WIDTH & STORESIZE_HEIGHT are the size of taken images. The final input size is STORESIZE_WIDTH * (STORESIZE_HEIGHT - CUT_SIZE).

2. Functions:

   • setInputImage(): Replace the image preprocess with your own. Now the preprocess is:

     data[j*image.rows*3+k*3+i] = (float(image.at<Vec3b>(j,k)[i])/255.0 - 0.5)*scale;

   • topKind(): It will use the results of softmax function and output the final kind, you only need it if you use classification model

   • runCV(): This function is independent of DPU usage, you can add your cv control in it.

   • If you use regression model, you don't need to use topKind() and softmax function, you should get the model output directly and add it in addSteer() or addSteer_Throttle instead of addCommand() and use it in run_steer() or run_steer_throttle() instead of run_command().

3. Output handle:

   The quantity of the model output is influenced by many factors. You can adjust the output to make the car run well. For example, if the car has a tendency to turn left, you can increase the steer value to make it normal.

Run the car

cd ~/Car
make run
./init.sh # instead run: insmod /home/xilinx/dpu.ko
./build/run n 0.5 # Usage of this exe: ./car c/n 0.5(run speed)

The first parameter means the run mode, n for only using neural network, c for using both neural network and computer vision functions. The second parameter is used unless the model outputs throttle value directly.

References

libn2cube API