Hyperparameter Selection for Lane Detection Applying Fully Convolutional Network with Attention Mechanism

Abstract

The Advanced Driver Assistance System (ADAS) significantly improves safety by offering several capabilities in smart navigation vehicles. An example of Advanced Driver Assistance Systems (ADAS) is lane detection. Previously, numerous studies have focused on detecting lanes for autonomous navigation vehicles. There is minimal research on the process of selecting hyperparameters for lane detection by combining attention processes with deep learning networks. This study attempts to optimize the network design of parameter setup in a deep learning model by introducing an attention mechanism to improve the accuracy of lane detection. An ideal network model for lane detection has been constructed using the Fully Convolutional Network (FCNs) model, and hyperparameter selection has been completed. The hyperparameter experiment involves adjusting parameters such as epochs, dropout rate, activation function, optimizer, max pooling, kernel filters, and loss function. Data collection involves collecting lane scenes with a camera sensor. The dataset is then inputted into the FCNs with an attention mechanism architecture, along with relevant labelling for the training phase. The model is validated using primary data and the KITTI benchmark lane dataset. Thus, an optimal combination resulting in high-accuracy performance is achieved. An epochs value of 10, dropout rate of 0.2, ReLU activation function, Adam optimizer, kernel filters of size (3,3) or (5,5), and MSE loss function result in high accuracy for the deep learning network model. The results show that the Fully Convolutional Networks (FCNs) using an attention mechanism reach above 96% performance accuracy.