-
目前关于车道线检测的数据集较多,包括CULane、Caltech Lanes、Mvirgo、TuSimple数据集等。
Mvirgo数据集是由Michael Virgo进行拍摄并标注,是专门为车道线检测算法的研发测试建立的数据集,总共包含12764张图片,数据集内包括直道、弯道、晴天、雨天、多云等多种场景的车道线数据,各种场景的数据分布如图6(a)所示,原始的图片和标签分别如图6(b)~(c)所示,对数据集进行处理后的图片和标签分别如图6(d)~(e)所示。
-
在确定了神经网络结构后,对TuSimple数据集,CULane数据集和Mvirgo数据集进行训练,并测试,测试结果如表1和图7所示,可以看到Mvirgo准确率最高,效果也最好。
Data sets TuSimple CULane Mvirgo TuSimple + Mvirgo Accuracy 97.5 93.21 98.23 97.8 Table 1. Accuracy of different data sets
由于3个数据各有特点,因此测试的准确率也各不相同,其中CULane数据集是由行车记录仪拍摄的照片,包含部分车身,车道线存在遮挡,因此测试效果最差,准确率也最低。TuSimple数据集训练效果也不好,主要原因是Tusimple数据集拍摄的图片较暗,车道线不明显,对神经网络的训练会产生一定的影响,导致检测准确率较低,对虚线的测量效果较差,虚警率高,测试结果如图7(b)所示。而Mvirgo数据集检测效果较好,主要原因是该数据集包含较多的场景,图片质量较高,缺点是该数据集数据量较少,并且只标注了当前道路的两条车道,对于一些特殊情况,无法处理,测试结果如图7(c)所示。另外,结合Tusimple数据集和Mvirgo数据集进行了训练,训练结果与只用Mvirgo数据集训练结果相似,测试结果如图7(d)所示,但对虚线的检测结果较差,因此选择了Mvirgo数据集。
-
笔者对比了现有的车道线检测网络,其中时间可能由于设备差异,存在区别,结果如表2所示,其中SegNetConvLSTM准确率最高,但是运行速度较慢,无法移植到海思平台做到实时检测,笔者提出的网络准确率与SegNetConvLSTM相当,但是速度方面具有明显的优势,可以移植到海思平台做到实时检测,在海思平台上的检测速度达到50 FPS,满足系统的要求。
Table 2. Comparison of different lane line detection network structures
-
考虑到驾驶场景的复杂性和多样性,选择多种场景的数据进行验证,测试结果如图8所示,结果表明该车道线检测方法在大多数场景中检测效果较好,但在人行横道、十字路口、车道线被完全遮挡的情况下表现不佳,针对这种情况,通过测距和设置阈值的方法,可以准确判断,依然能达到准确预警的效果。
Lane line detection method for embedded platform
doi: 10.3788/IRLA20210753
- Received Date: 2022-01-20
- Rev Recd Date: 2022-02-15
- Available Online: 2022-08-13
- Publish Date: 2022-08-05
-
Key words:
- lane line detection /
- embedded platform /
- deep learning /
- semantic segmentation /
- autonomous driving
Abstract: Lane line detection plays a pivotal role in autonomous driving and advanced assisted driving. However, traditional lane line detection technology was less robust, and most methods based on deep learning were more complex and difficult to embed platform real-time application. A lightweight lane line detection network for embedded platforms was proposed, which converts lane line detection into a semantic segmentation problem. The network draws on U-Net and Segnet network structures, and uses small-scale convolution and other lightweight components to design and calculate efficiently semantic segmentation network. Based on the detection of the lane line, calculate the distance between the vehicle and the lane line on both sides, as well as the curvature of the lane line, and give an early warning when the vehicle deviates from the lane line or the detection was abnormal. Finally, the entire system was transplanted to the HiSilicon platform. Experimental results show that the system has high detection accuracy and detection speed, the accuracy rate reaches 97.5%, the speed reaches 50 FPS, and meets real-time requirements.Therefore, the system can be used for real-time lane line detection, ranging, and distance measurement for embedded platforms. Curvature calculation and early warning.