Autonomous Vehicle Control for Lane and Vehicle Tracking by Using Deep Learning via Vision
Autonomous Vehicle Control for Lane and Vehicle Tracking by Using Deep Learning via Vision
复制标题
通过视觉使用深度学习进行车道和车辆跟踪的自主车辆控制
DOI：
10.1109/ceit.2018.8751764
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
Ozgur Koray Sahingoz
中科院分区：
文献类型：
--
作者：
Masum Celil Olgun;Zakir Baytar;Kadir Metin Akpolat;Ozgur Koray Sahingoz
基于摄像机的车道线检测和车辆跟踪算法是许多自主系统的关键技术之一。这些系统的导航过程主要集中在检测算法的输出。然而，用于车道检测的检测算法需要更多的预处理时间和计算工作量。它们还受到环境条件的影响，必须定期改进。在本文中，机器学习技术和计算机视觉算法被用于自动车辆控制场景的车道和车辆跟踪的任务。由于所使用的学习算法的性质，所提出的系统可以处理复杂的图像问题。我们实现算法的车辆可以自主执行以下任务：跟踪车道，跟随另一辆车，并在必要的条件下停止。为此，主要目的之一是通过使用学习算法的基于图像的车道跟踪方法。应用数据增强来创建数据集的多样性。讨论了该方法的应用。对于车道跟踪，首选基于NVIDIA PilotNet的卷积神经网络架构。为了检测物体和车辆，该系统在更快的基于区域的卷积神经网络（Faster R-CNN）上进行训练，以通过Haar级联分类器识别交通灯和停车标志。所有这些学习模型都在NVIDIA GTX 1070图形处理器（GPU）上进行训练，以减少训练时间。实验结果表明，所提出的系统给出了良好的效果，自主控制车辆的车道和车辆跟踪的视觉目的。
Camera-based lane detection and vehicle tracking algorithms are one of the keystones for many autonomous systems. The navigational process of those systems is mainly focused on the output of detection algorithms. However, detection algorithms for lane detection need more pre-processing time and computational effort. They are also affected by environmental conditions and must regularly be improved. In this paper machine learning techniques and computer vision algorithms are utilized for the tasks of the lane and vehicle tracking of an autonomous vehicle control scenario. With the nature of used learning algorithm, the proposed system can handle complex image problems. The vehicle, on which we implement our algorithms, can manage to carry out the following tasks autonomously; tracking the lanes, following another vehicle, and stopping in necessary conditions. For that, one of the primary purposes is image-based lane tracking methodology by using learning algorithms. Data augmentation is applied to create diversity for the dataset. Application in this methodology has been discussed. For lane tracking Convolutional Neural Network architecture which is based on NVIDIA’s PilotNet is preferred. For detecting objects and vehicles, the system is trained on the faster region-based convolutional neural network (Faster R-CNN) to identify traffic light and stop sign are by Haar Cascade Classifier. All these learning models are trained on NVIDIA GTX 1070 Graphics Processing Unit (GPU) to reduce training time. Experimental results showed that the proposed system gives a favorable result to autonomously control vehicles for lane and vehicle tracking purposes by vision.