Recently, the amount of research in the field of self-driving cars has grown significantly with autonomous vehicles having clocked in more than 10 million miles, providing a substantial amount of data for use in training and testing. The most complex part of training is the use of computer vision for feature extraction and object detection in real-time. Much relevant research has been done on improving the algorithms in the area of image segmentation. The proposed idea presents the use of Convoluted Neural Networks using Spatial Transformer Networks and lane detection in real time to increase the efficiency of autonomous vehicles. The depth of the neural network will help in training vehicles and during the testing phase, the vehicles will learn to make decisions based on the training data. In case of sudden changes to the environment, the vehicle will be able to make decisions quickly to prevent damage or danger to lives. Along with lane detection, a self-driving car must also be able to detect traffic signs. The proposed approach uses the Adam Optimizer which runs on top of the LeNet-5 architecture. The LeNet-5 architecture is analyzed and compared with the Feed Forward Neural Network approach. The accuracy of the LeNet-5 architecture was found to be 97% while the accuracy of the Feed Forward Neural Network was 94%.

近来，随着自动驾驶汽车的计时已超过1000万英里，无人驾驶汽车领域的研究量已显着增长，从而为训练和测试提供了大量数据。培训中最复杂的部分是使用计算机视觉实时进行特征提取和目标检测。在改进图像分割领域的算法方面，已经进行了许多相关的研究。提出的想法提出了使用通过空间变压器网络进行卷积神经网络和实时车道检测的方法，以提高自动驾驶汽车的效率。神经网络的深度将有助于训练车辆，在测试阶段，车辆将学习根据训练数据做出决策。如果环境突然发生变化，车辆将能够迅速做出决定，以防止造成人身伤害或生命危险。除车道检测外，自动驾驶汽车还必须能够检测交通标志。所提出的方法使用了运行在LeNet-5架构之上的Adam Optimizer。对LeNet-5架构进行了分析，并将其与前馈神经网络方法进行了比较。发现LeNet-5架构的准确性为97％，而前馈神经网络的准确性为94％。