1 how to transfer image to worldpoint
pointToWold, and use the measure distance to change world coordinate to vehicle coordinate.
2 how to process the cnn result with one times
pause(10) after detection and running.
3 lane detection
camera calibration: multiple angle of checkerboard to calibrate the camera. test different height of mouting, choose a good height.
vehilce platform: pwm 25%.
stanley controller
CNN: alexnet(90 images), cifar10net(40 images)
4camera
mounting height:33cm, angle:28 degrees.
5lane extraction
masking, kalman filter, inverse projection
kalman:Q=0.03, R=0.01. can reduce the effect of the light to the image.( varaiables estimate: ratio,coordinates)
phi=0.5*theta+e_fa (stanley)
resize the image to reduce the computation time.
max and min of steer angle is 0.8 and 0.2
drawbacks faced by image frame: depends on line, and affected by corner, curve.
move the vehilce and see the changing rule, then make a control decesion.
6 last group: group 4
use pointstoworld, because it is not easy to get the parameter 's' after calibration.
chanllege: kalman, it improved a little. but not changes the result. because the camera always detected the lane.
how to return the lane, if one of line is lost.
hough , peaks? change peak number.