Efficient Lane Detection Technique Based on Lightweight Attention Deep Neural Network

Table 3

Comparison with state-of-the-art methods on CULane dataset with IoU threshold = 0.5. Use PyTorch to test FPS and parameters at 3 ^∗ 288 ^∗ 800 resolution on NVIDIA RTX 2070s. Data with “” mark is obtained from the original text.

Category	Proportion	SCNN	ENet-SAD	Res18-VP	Res18-ultra	FastDraw	Proposed
Normal	27.74%	90.6	90.1	89.2	87.7	85.9	88.5
Crowded	23.39%	69.7	68.8	67.9	66.0	63.6	67.6
Night	20.27%	66.1	66.0	62.6	62.1	57.8	61.9
No line	11.73%	43.4	41.6	41.7	40.2	40.6	39.8
Shadow	2.68%	66.9	65.9	58.8	62.8	59.9	62.0
Arrow	2.57%	84.1	84.0	81.6	81.0	79.4	81.2
Dazzle light	1.40%	58.5	60.2	59.3	58.4	57.0	62.4
Curve	1.21%	64.4	65.7	60.8	57.9	65.2	56.9
Crossroad	9.00%	1990	1998	2919	1743	7013	2269
Total	—	71.6	70.8	69.1	68.4	—	68.8
FPS	—	8	48	42	285	—	259
Parameter	—	20.72	0.98	16.07	61.23	—	1.57