Algorithm 1 The second-stage model fine-training algorithm

Initialization:

1. Collect the image sets under various scenarios; 2. First-stage model training hlane_1;

3. Set the confidence threshold T; 4. Set the threshold growth factor ζ;

5. Divide images into multiple sets M, of which each contains the batch size number of images

Training process:

For batch in M:

for Xi in batch:

1. Obtain the coordinates, width, and height (x, y, w, h), and confidence ξ using the model hlane_1

if T ≤ ξi:

(1) Set a new area of the bounding box as (x, y, w + 2δ, h + 2δ)

(2) Determine the lane edge using the adaptive threshold detection algorithm based on the Canny

(3) Binarize the area and determine the connected domain of the lane

(4) Determine a new bounding box (x’, y’, w’, h’) of the lane

else if T/4 ≤ ξi < T:

(1) Set a new area of the bounding box as (x, y, w + 2δ, h + 2δ)

(2) Set the pixel value in the new area to 0

end

end

2. Obtain the processed image sets $\mathit{X}=\{X_i^{\prime },X_{i+1}^{\prime },\dots ,X_{i+batchsize-1}^{\prime }\}$

3. Make T = T + ζ, retrain hlane_1

end