Mapping Road Lanes Using Laser Remission and Deep Neural Networks

Abstract

We propose the use of deep neural networks (DNN) for solving the problem of inferring the position and relevant properties of lanes of urban roads with poor or absent horizontal signalization, in order to allow the operation of autonomous cars in such situations. We take a segmentation approach to the problem and use the Efficient Neural Network (ENet) DNN for segmenting LiDAR remission grid maps into road maps. We represent road maps using what we called road grid maps. Road grid maps are square matrixes and each element of these matrixes represents a small square region of real-world space. The value of each element is a code associated with the semantics of the road map. Our road grid maps contain all information about the roads' lanes required for building the Road Definition Data Files (RDDFs) that are necessary for the operation of our autonomous car, IARA (Intelligent Autonomous Robotic Automobile). We have built a dataset of tens of kilometers of manually marked road lanes and used part of it to train ENet to segment road grid maps from remission grid maps. After being trained, ENet achieved an average segmentation accuracy of 83.7%. We have tested the use of inferred road grid maps in the real world using IARA on a stretch of 3.7 km of urban roads and it has shown performance equivalent to that of the previous IARA's subsystem that uses a manually generated RDDF.

A Specialized Database for Autonomous Vehicles Based on the KITTI Vision Benchmark

“…Examples of the above can be found in [22], in which a comparison is made between a U-Net and a fully convolutional network (FCN) for the road segmentation task in 2D images of aerial views, where the metrics of each neural network with different image sizes stand out. In the same way, the efficient neural network (ENet) can be used to infer the position and properties of the lines of a highway by segmenting roads in 2D images of aerial maps and generating a grid map that identifies the lateral and central lines of the road, seeking to support autonomous vehicle management with the generation of this information [23]. Another example of this is found in [24], where researchers used 3D scenes acquired with radar to generate 2D top-view maps that they used as input for the fully convolutional neural network, SegNet, and U-Net networks, in an effort to segment the streets, cars, edges, and fences as the output of the network in the 2D top-view plan form.…”

“…In the state-of-the-art databases for autonomous driving that have scenes generated from point clouds [22][23][24][25], the data therein are three-dimensional. Therefore, the detection labels have three location values.…”

Extending Maps with Semantic and Contextual Object Information for Robot Navigation: a Learning-Based Framework Using Visual and Depth Cues

“…The RGB-D sequences used two different RGB-D sensing cameras: Microsoft Kinect (sequence1-Kinect) and Orbbec Astra (sequence2-Astra and sequence3-Astra). Further details of time duration, data statistics and information parsers is given in the project page 3 .…”

“…In this context, embedding a higher level of scene understanding to identify particular objects of interest (including people), as well as to localize them, would greatly benefit intelligent agents to perform effective visual navigation, perception and manipulation tasks. Notably, this is a desired capability in human-robot interaction or in autonomous robot navigation tasks in daily-life scenes since it can provide "situation awareness" by distinguishing dynamic entities (e.g., humans, vehicles) from static ones (e.g., door, bench) [2,3], or to recognize unsafe situations. This competence is also instrumental in the development of personal assistant robots, which need to deal with different objects of interest for guiding visually impaired people to cross a door, to find a bench, or a water fountain.…”

Self-Driving Cars: A Survey

“…However, a grid representation might require a wasteful use of memory space and processing time, since usually most of the environment where self-driving cars operate is not composed of roads, but buildings, free space, etc. Carneiro et al (2018) proposed a metric road map (Figure 5), a grid map, where each 0.2 m × 0.2 m-cell contains a code that, when nonzero, indicates that the cell belongs to a lane. Codes ranging from 1 to 16 represent relative distances from a cell to the center of the lane, or the type of the different possible lane markings (broken, solid, or none) present in the road.…”