notesum.ai

Published at November 25

Monocular Lane Detection Based on Deep Learning: A Survey

cs.CV

Released Date: November 25, 2024

Authors: Xin He¹, Haiyun Guo¹, Kuan Zhu¹, Bingke Zhu¹, Xu Zhao¹, Jianwu Fang², Jinqiao Wang¹

Aff.: ¹School of Artificial Intelligence, University of Chinese Academy of Sciences; ²Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University

Arxiv: http://arxiv.org/abs/2411.16316v1


Method	Backbone	Input Size	FPS	Output	Post-processing

2D Lane Detection Methods (Section III)

SCNN [36]	ResNet18	288 $\times$ 800	14	Multi-classes semantic segmentation mask.	Vectorization
SAD [56]	ERFNet	208 $\times$ 976	92	Multi-classes semantic segmentation mask.	Vectorization
RESA [16]	ResNet18	288 $\times$ 800	68	Multi-classes semantic segmentation mask.	Vectorization
LaneAF [66]	ERFNet	288 $\times$ 832	63	Binary segmentation mask and affinity vector fields.	Clustering $\&$ Vectorization
UFLD [68]	ResNet18	288 $\times$ 800	358	Multi classification probability of grid for each row.	Vectorization
LSTR [82]	ResNet18	288 $\times$ 800	133	Each cubic polynomial’s classification probability and coefficient values.	None
LaneATT [80]	ResNet18	360 $\times$ 640	194	Lane’s classification probability and equidistant point coordinates.	NMS
CondLaneNet [17]	ResNet18	320 $\times$ 800	219	Multi classification probability of grid for each row.	Vectorization
BézierLaneNet [20]	ResNet18	288 $\times$ 800	244	Each cubic Bézier curve‘s classification probability and control point coordinates.	None
GANet [18]	ResNet18	320 $\times$ 800	106	Binary keypoint heatmap and offset map.	Clustering
CLRNet [18]	ResNet18	320 $\times$ 800	104	Lane’s classification probability and equidistant point coordinates.	NMS
CondLSTR [18]	ResNet18	320 $\times$ 800	47	Keypoint heatmap and offset map for each instance.	Refinement
ADNet [18]	ResNet18	320 $\times$ 800	109	Lane’s classification probability and equidistant point coordinates.	NMS

3D Lane Detection Methods (Section IV)

3D-LaneNet [22]	VGG16	360 $\times$ 480	118	BEV lane’s classification probability and equidistant point coordinates, 3D lane heights.	NMS
Gen-LaneNet [23]	ERFNet	360 $\times$ 480	24	BEV lane’s classification probability and equidistant point coordinates, 3D lane heights.	NMS
PersFormer [24]	EfficientNetB7	360 $\times$ 480	19	BEV lane’s classification probability and equidistant point coordinates, 3D lane heights.	NMS
PersFormer [24]	EfficientNetB7	720 $\times$ 960	12	BEV lane’s classification probability and equidistant point coordinates, 3D lane heights.	NMS
Anchor3DLane [25]	ResNet18	360 $\times$ 480	75	3D lane’s classification probability and equidistant point coordinates.	NMS
Anchor3DLane $\dagger$ [25]	ResNet50	720 $\times$ 960	18	3D lane’s classification probability and equidistant point coordinates.	NMS
BEV-LaneDet [26]	ResNet34	576 $\times$ 1024	83	BEV lane’s equidistant point coordinates and instance embedding, 3D lane heights.	Clustering
SPG3DLane [121]	EfficientNetB7	720 $\times$ 960	13	BEV lane’s classification probability and equidistant point coordinates, 3D lane heights.	NMS
LATR [27]	ResNet50	720 $\times$ 960	14	3D lane classification probability and equidistant point coordinates.	None
LATR-Lite [27]	ResNet50	720 $\times$ 960	22	3D lane classification probability and equidistant point coordinates.	None