Comparing of Some Convolutional Neural Network (CNN) Architectures for Lane Detection

Abstract

Advanced driver assistance functions help us prevent the human-based accidents and reduce the damage and costs. One of the most important functions is the lane keeping assist which keeps the car safely in its lane by preventing careless lane changes. Therefore, many researches focused on the lane detection using an onboard camera on the car as a cost-effective sensor solution and used conventional computer vision techniques. Even though these techniques provided successful outputs regarding lane detection, they were time-consuming and required hand-crafted stuff in scenario-based parameter tuning. Deep learning-based techniques have been used in lane detection in the last decade. More successful results were obtained with fewer parameter tuning and hand-crafted things. The most popular deep learning method for lane detection is convolutional neural networks (CNN). In this study, some reputed CNN architectures were used as a basis for developing a deep neural network. This network outputs were the lane line coefficients to fit a second order polynomial. In the experiments, the developed network was investigated by comparing the performance of the CNN architectures. The results showed that the deeper architectures with bigger batch size are stronger than the shallow ones.

Analyzing Performance Effects of Neural Networks Applied to Lane Recognition under Various Environmental Driving Conditions

“…The selected images were adjusted to the network architectures' input layer size. This change was made to avoid the high computational cost and reduce the training time involved in processing images with huge resolutions [32]. Similarly, by reducing the images' size, the networks can see key features in the initial layers, which would otherwise be learned at the network's end [43].…”