Road lane segmentation using deconvolutional neural network

Abstract

Lane departure warning (LDW) system attached to modern vehicles is responsible for lowering car accident caused by inappropriate lane changing behaviour. However the success of LDW system depends on how well it define and segment the drivable ego lane. As the development of deep learning methods, the expensive light detection and ranging (LIDAR) guided system is now replaced by analysis of digital images captured by low-cost camera. Numerous method has been applied to address this problem. However, most approach only focusing on achieving segmentation accuracy, while in the real implementation of LDW, computational time is also an importance metric. This research focuses on utilizing deconvolutional neural network to generate accurate road lane segmentation in a realtime fashion. Feature maps from the input image is learned to form a representation. The use of convolution and pooling layer to build the feature map resulting in spatially smaller feature map. Deconvolution and unpooling layer then applied to the feature map to reconstruct it back to its input size. The method used in this research resulting a 98.38% pixel level accuracy and able to predict a single input frame in 28 ms, enabling realtime prediction which is essential for a LDW system.