Lane Detection Using CNN-LSTM with Curve Fitting for Autonomous Driving

The efficient and accurate detection of lanes and the extraction of their key features are critical to autonomous driving. In this paper, a lane detection method that combines convolutional neural networks (CNN) and long-short-time memory neural networks (LSTM) is proposed to extract key features of lanes with great rapidity and accuracy. The main process is as follows: （ 1 ） The video is processed using a featurebased image processing method to extract key information of the lanes which is stored as a label. (2) The CNN model and the CNN-LSTM model are established respectively. （ 3 ） Training and testing are operated on above-mentioned models using the images and labels obtained in step(1). （ 4 ） Multi-platform verification of trained models is operated with entirely new videos. The results show that the detection rates of CNN model on training data and verification data are 94.9% and 91.2%, respectively, and the processing speed reaches up to 46.2 milliseconds per frame and its time consumption is only 5.59% of the traditional processing method; the detection rates of CNN-LSTM model are respectively 97.6% and 94.4%, and the processing speed achieves 54.7 milliseconds per frame which consumes only 6.61% time of the traditional method, and it also shows great performance on the micro platform.