PaperlayerView exact revision on arXiv

Source-linked AI summary

Mapping Road Lanes Using Laser Remission and Deep Neural Networks

Raphael V. Carneiro, Rafael C. Nascimento, Rânik Guidolini, Vinicius B. Cardoso, Thiago Oliveira-Santos, Claudine Badue, Alberto F. De Souza

arXiv:1804.10662v1cs.ROcs.NE

TL;DR

The paper addresses how autonomous cars can infer urban road lanes when horizontal signalization is poor or absent. It represents lane semantics in road grid maps and uses ENet to segment LiDAR remission grid maps into those maps. ENet achieved 83.7% average segmentation accuracy, and IARA matched the performance of a manually generated-RDDF subsystem over 3.7 km of urban roads.

  • Problem

    Autonomous cars need lane information to operate on urban roads, but horizontal signalization can be poor or absent.

    Source: §I · p. 1

  • Method

    The method uses ENet to segment LiDAR remission grid maps into semantic road grid maps from which RDDFs can be extracted.

    Source: §I · p. 1, A DNN for Semantic Segmentation of… · p. 3, §VI · p. 7

  • Results

    IARA showed performance equivalent to its previous manually generated-RDDF subsystem on 3.7 km of urban roads, while ENet achieved 83.7% average segmentation accuracy.

    Source: §I · p. 2, §VI · p. 7

  • Takeaways & Limitations

    The inferred road grid maps supported autonomous operation with IARA at performance equivalent to manually generated RDDFs in the tested urban-road setting.

    Source: §VI · p. 7

Abstract

from arXiv · show

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.

I. INTRODUCTION

The paper addresses lane inference for autonomous driving on urban roads where horizontal markings are poor or absent. It uses ENet to segment LiDAR remission maps into road maps and tests whether the resulting maps support autonomous operation.

  • ENet segments LiDAR remission grid maps into road maps to infer lane positions and properties when horizontal signalization is poor or absent.The approach targets autonomous cars that cannot rely on road markings alone.Source: §I · p. 1
  • 83.7% average segmentation accuracy was achieved on a dataset containing tens of kilometers of manually marked road lanes.The authors attribute part of the apparent error to difficult manual annotation and note that the DNN can infer lanes absent from the ground truth.Source: §I · pp. 1–2
  • 3.7 km of urban-road testing with IARA produced performance equivalent to its previous subsystem using manually generated RDDFs.The DNN-generated road maps were converted automatically into RDDFs for vehicle operation.Source: §I · p. 2
  • IARA is an autonomous car developed at Universidade Federal do Espírito Santo that uses multiple sensors and previously required precisely annotated RDDFs.RDDFs contain regularly spaced world waypoints and associated annotations such as road width and maximum speed.Source: §I · p. 2, Figure 2 · p. 2

II. RELATED WORK

The related work covers image-, LiDAR-, and multimodal approaches to road and lane detection, while distinguishing this paper's road-map output and real-vehicle evaluation. The proposed representation encodes lane semantics in grid cells and uses ENet for semantic segmentation.

Prior approaches

  • Prior methods detect lanes or roads from images, LiDAR, or fused data, but several do not produce road maps or lack evaluation on a real autonomous car.The cited approaches include temporal image processing, pixel-level segmentation, LiDAR-image fusion, and reflectivity-based lane detection.Source: §II · pp. 2–3

Road-grid representation

Segmentation method

  • ENet uses an encoder-decoder architecture to segment road grid maps from remission grid maps, with an encoder for classification and a decoder for upsampling.The architecture was modified for faster operation without significant degradation of segmentation performance and is trained in two stages.Source: A DNN for Semantic Segmentation of… · p. 3
  • During testing, ENet assigns each road-grid cell the most active class from its corresponding softmax classifier.The network receives a crop of the remission grid map and outputs an inferred road grid map.Source: A DNN for Semantic Segmentation of… · p. 3

III. THE USE OF ROAD GRID MAPS IN AUTONOMOUS CARS

Road grid maps encode lane semantics for autonomous-car operation and can be used to derive RDDFs for IARA.

  • Road grid maps can inform lane-center distance, overtaking-side availability, and road-network structure for autonomous cars.Source: §III · p. 3
  • IARA uses road grid maps to build Road Definition Data Files for autonomous operation.Source: §III · p. 4
  • RDDFs contain waypoints, recommended orientations, and road annotations such as speed limits, speed bumps, and crosswalks.Source: §III · p. 4

A. The IARA Software System

IARA’s autonomy is organized as cooperating software modules that exchange messages, retrieve and maintain maps, and process mission and obstacle information.

B. Computing RDDFs from Road Grid Maps

The system converts road grid maps into cropped RDDFs by generating lane-centered waypoints ahead and behind IARA, then smoothing them.

IV. EXPERIMENTAL METHODOLOGY

The real-world evaluation uses IARA, a self-developed autonomous car whose hardware and software support autonomous operation.

  • The study evaluates ENet-segmented road lanes in the real world using the IARA autonomous car.Source: §IV · p. 5
  • IARA is based on a Ford Escape Hybrid modified for electronic vehicle control, odometry, computing, and sensor integration.Source: §IV · p. 5
  • Its sensors include a Velodyne HDL-32E LiDAR, dual RTK GPS, IMU, stereo cameras, and two high-performance computers.Source: §IV · p. 5
  • IARA’s software modules exchange messages through publish-subscribe communication in the UFES CARMEN framework.Source: §IV · p. 5

A. Datasets

The dataset combines manually marked lane maps from a 3.7 km UFES ring road and 32.4 km of highways. UFES samples were augmented through rotations and translations, while the Highway dataset was reserved for testing generalization.

  • The dataset contains manually marked road lanes spanning 3.7 km of the UFES ring road and 32.4 km of highways.Source: A. Datasets · pp. 5–6
  • UFES ground truth was created by drawing lane-center polylines, fitting cubic Bezier curves, and assigning standardized 3.2 m lane widths and color codes.Source: A. Datasets · pp. 5–6
  • Each UFES crop was expanded into 168 samples using 24 rotations and 7 translations, producing 110,544 crop pairs.Source: A. Datasets · p. 6
  • The Highway dataset contains 3,556 unaugmented 24 m × 24 m crops sampled about 12 m apart and was used only to evaluate generalization to unseen surroundings.Source: A. Datasets · p. 6

B. ENet DNN

The system uses a modified ENet to map 120 × 120 grayscale remission-grid crops to road-grid-map ground truth. Performance is assessed with class average accuracy alongside navigation-oriented criteria.

  • The modified ENet receives 120 × 120 grayscale remission-grid images and predicts 120 × 120 road-grid-map ground truth images.Source: B. ENet DNN · p. 6
  • Training used the UFES training dataset, while evaluation used both the UFES test dataset and the Highway dataset.Source: B. ENet DNN · p. 6
  • Performance was measured by class average accuracy, defined as correctly inferred classes divided by evaluated map cells.Source: B. ENet DNN · p. 6
  • Navigation quality was also judged by lane-center keeping and avoidance of abrupt steering movements.Source: B. ENet DNN · p. 6

V. EXPERIMENTS

ENet reached about 78% accuracy during training, 83.7% on the UFES test set, and 64.1% on unseen highways. Road-grid maps inferred on the UFES ring road supported autonomous IARA navigation without human intervention, matching the manually annotated baseline.

  • Training accuracy plateaued at about 78% after increases associated with early learning-rate reductions.Source: §V · p. 6, Figure 8 · p. 6
  • 83.7% average accuracy was achieved on the UFES test dataset, exceeding the training accuracy measured during training.Source: §V · p. 6, Figure 9 · p. 6
  • 64.1% average accuracy was achieved on the Highway dataset, where the test region was unseen during training and came from different surroundings.Source: §V · pp. 6–7, Figure 11 · p. 7
  • IARA achieved navigation performance equivalent to the manually annotated RDDF baseline using an RDDF extracted from ENet-inferred road maps, with no human intervention.Source: §V · p. 7
  • The application-oriented evaluation emphasizes whether inferred maps preserve lane shape and center, rather than relying on a universally optimal segmentation measure.Source: §V · p. 7

VI. CONCLUSIONS AND FUTURE WORK

The paper combines ENet semantic segmentation of LiDAR remission grid maps with road-grid-map processing to support autonomous driving on urban roads with poor or absent horizontal signalization. On 3.7 km of urban roads, the approach matched the performance of manually generated RDDFs, while future work targets broader architectures, datasets, and road-network inference.

Approach

  • ENet segmented LiDAR remission grid maps into road maps, from which simple algorithms extracted RDDFs to guide autonomous cars.Road grid maps encode the semantics of small real-world regions and contain the lane information needed to build RDDFs.Source: §VI · p. 7

Results

  • ENet achieved 83.7% average segmentation accuracy on a dataset built from manually marked tens of kilometers of roads.The reported accuracy also reflects imperfect manual ground truth and cases where the DNN inferred existing lanes absent from annotations.Source: §VI · p. 7
  • Performance on 3.7 km of urban roads was equivalent to a system using manually generated RDDFs.The authors suggest that 83.7% segmentation accuracy was sufficient for proper autonomous operation with their approach.Source: §VI · p. 7

Future work

  • Future work includes tuning and replacing the DNN architecture, training and testing on the complete dataset, and using transfer learning to reduce repeated labeling.The authors also plan to infer full road networks as graphs and investigate online mapping tools for high-level path planning.Source: §VI · p. 7
Loading 1804.10662v1…