MLP-based Efficient Convolutional Neural Network for Lane Detection

Lane detection is an important and fundamental task in autonomous driving. Modern convolutional neural network (CNN) methods have achieved high performance lane detection; however, the intrinsic locality of convolution operations makes these limited effectively modeling long-range dependencies that are vital to capture global information lanes. Additionally, numerous result considerable computational cost for complexity. To overcome difficulties, we propose efficient method by combining CNN with a multilayer perceptron (MLP). First, improved bottleneck-1D layer used replace standard overall reduce parameters while applying hybrid dilated (HDC) better multiscale information. Second, construct MLP block latent space The projects tokenized features from spatial locations channels, then, they fused together obtain representation, which each output pixel related input pixel. introduction further decreases complexity proposed architecture more detection. Experimental results on two challenging datasets (CULane, Tusimple) demonstrate our can achieve higher efficiency maintaining decent compared other state-of-the-art methods. Furthermore, this study indicates integrating representation capacity local prior effective potential perspective