A new lightweight convolutional neural network model for detecting drivable road regions

Abstract

Nowadays, due to the rapid increase in the number of autonomous vehicles on the market, the safe navigation of these vehicles in drivable road areas has become extremely important. One of the most crucial factors in ensuring safe navigation is addressing the detection of drivable road areas as a task of semantic segmentation. Considering that autonomous vehicles are modular, the algorithm to perform this task must have the optimum trade-off in terms of lightweight, computational complexity, and segmentation accuracy. In this study, RoNet, a new model based on convolutional neural networks that provides an optimum trade-off for the detection of drivable road regions, was designed and proposed. The standard convolution types for the encoder and decoder bottleneck module of the RoNet model, as well as the spatial edge attention mechanism, have been optimized by developing asymmetric convolution types using asymmetric atrous convolution, asymmetric convolution types using Prewitt and Sobel kernels. Spatial edge attention mechanism is designed to reduce the loss of detailed information in small-resolution feature maps. In experimental tests performed with CamVid and FUVid datasets, RoNet achieved a better trade-off in terms of segmentation accuracy, number of parameters, and computational complexity compared to other state-of-the-art methods.