Unstructured road segmentation algorithm based on improved BiSeNet

Liang SONG; Yuhai GU; Wentian SHI

doi:10.5768/JAO202344.0302003

Journal of Applied Optics, Volume. 44, Issue 3, 556(2023)

Unstructured road segmentation algorithm based on improved BiSeNet

Liang SONG1... Yuhai GU1,* and Wentian SHI2 |Show fewer author(s)

¹Beijing Key Laboratory of Measurement and Control of Mechanical and Electrical System Technology, Beijing Information Science and Technology University, Beijing 100192, China

²School of Materials Science and Mechanical Engineering, Beijing Technology and Business University, Beijing 100148, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(13)

Fig. 1. Structure diagram of BiseNet network

Download full size

View in Article

Fig. 2. Schematic of depthwise separable convolution

Download full size

View in Article

Fig. 3. Schematic of improved spatial path

Download full size

View in Article

Fig. 4. Structure diagram of Bneck

Download full size

View in Article

Fig. 5. Schematic of improved context path

Download full size

View in Article

Fig. 6. Schematic of channel attention feature fusion module

Download full size

View in Article

Fig. 7. Schematic of category reclassification of IDD dataset

Download full size

View in Article

Fig. 8. Preprocessing results graph

Download full size

View in Article

Fig. 9. Variation curves of F1-score and loss during training

Download full size

View in Article

Fig. 10. Comparison chart of prediction effect

Download full size

View in Article

Table 1. MobileNetV3-large feature extraction network after adjustment

View table

View in Article

Table 1. MobileNetV3-large feature extraction network after adjustment

Input（H²×C）	操作	n	SE	激活函数	s	out
表中：H、C表示特征的边长、通道数；Conv2d表示标准卷积；n表示需进行操作的次数；SE表示是否引入SE注意力；s表示步长；out表示输出通道数。
512² × 3	Conv2d	1	-	HS	2	16
256² × 16	Bneck,3 × 3	1	-	ReLU	1	16
256² × 16	Bneck,3 × 3	1	-	ReLU	2	24
128² × 24	Bneck,3 × 3	1	-	ReLU	1	24
128² × 24	Bneck,5 × 5	1	√	ReLU	2	40
64² × 40	Bneck,5 × 5	2	√	ReLU	1	40
64² × 40	Bneck,3 × 3	1	-	HS	2	80
32² × 80	Bneck,3 × 3	3	-	HS	1	80
32² × 80	Bneck,3 × 3	2	√	HS	1	112
32² × 112	Bneck,5 × 5	1	√	HS	2	160
16² × 160	Bneck,5 × 5	2	√	HS	1	160

Table 2. Model performance with different improvement measures

View table

View in Article

Table 2. Model performance with different improvement measures

模型	MIOU/%	F1-score/%	Params
Conv+Xception+FFM	89.83	92.66	1.54×10⁶
Conv+ResNet+FFM	91.93	94.32	1.29×10⁷
Conv+MobileNetv3+FFM	92.43	94.65	1.16×10⁶
DSConv+MobileNetv3+FFM	92.68	94.80	1.09×10⁶
DSConv+MobileNetv3+CAFFM	95.40	96.74	1.11×10⁶

Table 3. Performance comparison of different network models

View table

View in Article

Table 3. Performance comparison of different network models

模型	F1-Score/%	PA/%	MIOU/%	Params	Speed/f·s⁻¹
U-Net	91.56	95.67	89.38	1.94×10⁶	26.17
DeepLabV3+	95.85	97.46	94.23	3.87×10⁷	15.21
FCN-32S	92.54	96.18	90.56	5.44×10⁷	28.63
BiSeNet-resnet	94.32	96.27	91.93	4.92×10⁷	23.57
BiSeNet-xception	92.66	95.31	89.83	1.54×10⁶	30.96
本文	96.74	97.78	95.40	1.11×10⁶	36.79