Lane Instance Segmentation Algorithm Based on Convolutional Neural Network

Fig. 7. Test results of different scenarios. (a) Three-lane (straight); (b) four-lane (curve); (c) vehicle occlusion environment; (d) light and shadow occlusion environment

Download full size

Table 1. Predictive branch network for the number of lanes

View table

Table 1. Predictive branch network for the number of lanes

Structure	Convolution kernel number	Convolution kernel size/stride	Output size
Separable convolution+BN+ReLU	16	3×3	64×128
Separable convolution+BN+ReLU	16	3×3	64×128
Maxpool		2×2/2	32×64
Separable convolution+BN+ReLU	32	3×3	32×64
Separable convolution+BN+ReLU	32	3×3	32×64
Maxpool		2×2/2	16×32
Separable convolution+BN+ReLU	64	3×3	16×32
Separable convolution+BN+ReLU	64	3×3	16×32
Maxpool		2×2/2	8×16
Separable convolution+BN+ReLU	128	3×3	8×16
Separable convolution+BN+ReLU	128	3×3	8×16
Maxpool		2×2/2	4×8
Global_avgpool		4×8	1×1
Separable convolution+BN+ReLU		1×1	X_Max+1

Table 2. Accuracies of different algorithms unit: %
View table
Table 2. Accuracies of different algorithms unit: %
Algorithm Accuracy FP FN
Ref.[7] 96.53 6.17 1.80
Ref.[16] 96.50 8.51 2.69
Ref.[8]1 96.40 7.80 2.44
Ref.[17] 95.24 11.94 6.20
Ref.[8]2 96.20 23.58 3.62
Ours 96.23 8.34 7.29

Table 3. Running time of different algorithms unit: ms
View table
Table 3. Running time of different algorithms unit: ms
Algorithm LaneNet Ours
Module Time Module Time
Network inference E-Net 80 DeepLabV3+ with DenseASPP 76
Cluster Mean shift 85 K-means with lane number branch 49
Lane fitting Least squares 4 Least squares 4
Sum 169 129

Tools

Get Citation

Copy Citation Text

Su Zhou, Di Wu, Jie Jin. Lane Instance Segmentation Algorithm Based on Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2021, 58(8): 0815007

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites