| 引用本文: | 张 绪,林玉娥,王 慧,梁兴柱.基于双分支分解卷积的夜间车道线检测方法(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2026，43（2）：125-132 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 目的 车道线检测作为自动驾驶系统的核心功能之一,对于确保车辆安全行驶至关重要。 然而,当前基于分
割的检测方法在速度和性能之间存在明显不平衡,尤其在夜间场景下更为突出。 针对这一挑战,提出一种名为
DecoLaneNet 的夜间车道线检测模型。 方法 首先,设计一种高效的特征提取模块,采用分解卷积构建具有不同大小
感受野的双分支分解卷积残差结构,在尽可能保证性能的情况下大幅减少模型的参数量,更准确地提取车道线特
征;接着,运用通道重排技术和扩张卷积弥补性能损失,并增加感受野;最后,引入双分支上采样模块进行特征解
码,显著提升了模型的分割精度。 结果 在夜间场景数据集 Night 和包含多种交通场景的车道检测数据集 CULane
上进行了广泛评估,DecoLaneNet 的 F1 评分,在 Night 数据集上达到了 74. 7%,在 CULane 数据集上达到了 71. 5%。
值得一提的是,尽管模型参数仅有 1. 94 M,但在 Jetson 开发板上,其帧率(FPS)仍能达到 62. 5。 结论 实验结果表
明,DecoLaneNet 不仅在夜间场景下表现出优异的性能,在应对其他复杂场景时同样表现出色。 此外,在部署到嵌
入式设备上时,其性能与效率仍然优于其他模型,显示出了出色的潜力与可行性。 |
| 关键词: :车道线检测 通道重排 分解卷积 模型部署 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Nighttime Lane Detection Method Based on Dual-branch Decomposed Convolution |
|
ZHANG Xu LIN Yu?? e WANG Hui LIANG Xingzhu
|
|
School of Computer Science and Engineering Anhui University of Science and Technology Huainan 232001 Anhui
China
|
| Abstract: |
| Objective Lane line detection as one of the core functions of autonomous driving systems is crucial for
ensuring the safe driving of vehicles. However current segmentation-based detection methods show a significant imbalance
between speed and performance which is particularly prominent in nighttime scenarios. To address this challenge this
paper proposes a nighttime lane line detection model called DecoLaneNet. Methods First an efficient feature extraction
module was designed. In this module a dual-branch decomposed convolution residual structure with different-sized
receptive fields was constructed using decomposed convolutions which significantly reduced the number of model
parameters while ensuring performance as much as possible and more accurately extracted lane line features. Then
channel rearrangement technology and dilated convolutions were used to make up for performance losses and increase the
receptive field. Finally a dual-branch upsampling module was introduced for feature decoding which significantly
improved the model?? s segmentation accuracy. Results Extensive evaluations were conducted on the nighttime scenario dataset Night and the lane detection dataset CULane containing various traffic scenarios. The F1
score of DecoLaneNet
reached 74. 7% on the Night dataset and 71. 5% on the CULane dataset. Remarkably despite having only 1. 94 M
parameters the model maintained a speed of 62. 5 FPS on the Jetson development board. Conclusion The experimental
results show that DecoLaneNet not only performs excellently in nighttime scenarios but also shows good performance in
dealing with other complex scenarios. Moreover when the model is deployed on embedded devices its performance and
efficiency are still better than those of other models showing excellent potential and feasibility. |
| Key words: lane detection channel shuffle decomposed convolution model deployment |
