Automatic detection of zebra crossings from mobile LiDAR data

Highlights

• Road segmentation using curvature analysis.

• LiDAR to image conversion of road data.

• Implementation of Generalized Hough Transform to detect zebra crossings.

Abstract

An algorithm for the automatic detection of zebra crossings from mobile LiDAR data is developed and tested to be applied for road management purposes. The algorithm consists of several subsequent processes starting with road segmentation by performing a curvature analysis for each laser cycle. Then, intensity images are created from the point cloud using rasterization techniques, in order to detect zebra crossing using the Standard Hough Transform and logical constrains. To optimize the results, image processing algorithms are applied to the intensity images from the point cloud. These algorithms include binarization to separate the painting area from the rest of the pavement, median filtering to avoid noisy points, and mathematical morphology to fill the gaps between the pixels in the border of white marks. Once the road marking is detected, its position is calculated. This information is valuable for inventorying purposes of road managers that use Geographic Information Systems.

The performance of the algorithm has been evaluated over several mobile LiDAR strips accounting for a total of 30 zebra crossings. That test showed a completeness of 83%. Non-detected marks mainly come from painting deterioration of the zebra crossing or by occlusions in the point cloud produced by other vehicles on the road.

Introduction

Cities are increasingly large and complex areas that require integrated technologies for an effective management and to ensure productivity, continued economic growth and environmental sustainability. The city management is increasingly supported by information technologies, leading to paradigms such as smart cities, where decision-makers, companies, and citizens are continuously interconnected. One of the computational tools that support smart cities management is the Geographic Information Systems (GIS). They organize information in layers with georeferenced data relevant to the city management. Typical GIS layers for urban management consist of urban elements such as horizontal and vertical signs, bus shelters, rubbish containers, lamps, banks, gardens, fountains, etc.

One of the most used spatial data acquisition technologies comprises the mobile LiDAR systems that achieve geo-reference point clouds and imagery in an accurate and productive manner. Mobile LiDAR systems acquire massive information, which must be organized in order to extract data of interest for geospatial analysis [1], [2], [3]. For example, a point cloud with several thousands of 3D points that describe the geometry of a traffic light is not very useful, although its geographic position of the urban object, height, manufacturer, type of luminaire, and maintenance events are aspects of interest for the infrastructure manager. Then, it can be affirmed that the users of LiDAR data more and more demand the development of algorithms to automatically extract useful information from large point clouds. This fact is corroborated by the intense work reported in the literature in recent years about automatic detection and classification of urban objects from LiDAR data.

Jaakkola et al. [4] develop algorithms for the detection of road curbstones using gradients in height plane. Hernández and Marcotegui [5] developed a methodology for the segmentation of the pavement contour using a quasi-flat zone algorithm. Pu et al. [6] present a framework for structure recognition from mobile laser scanner point cloud. It performs a rough classification of elements into three main categories (ground surface, objects on ground, and objects off ground). Using characteristics from the point cloud like size, shape, orientation and topological relationships, objects on ground are more detailed classified in signs, trees, building walls, and barriers. Mc. Ehinney et al. [7] design an algorithm for extracting the road edge using LiDAR and navigation data. Kumar et al. [8] developed an algorithm for extracting road edges using mobile LiDAR data. The algorithm is based on the snake model. It uses parameters selected empirically and fixed for all the road sections. The snake model is initialized based on the navigation information from the navigation system. Puente et al. [9] detect luminaires on a road tunnel using the RGB images registered over the point cloud of the tunnel. They perform an automated inventory in a productive and robust manner. Wang et al. [10] estimate the excavation volume in road widening from mobile LiDAR data. Normals of the points and slopes are used to separate road from off-road points. The left and right sides of the road points are sliced up to a distance of 4 m perpendicular to the roadside. Summing each sliced volume permits the calculation of the required excavation for a road widening. Cabo et al. [11] show an algorithm for the automatic extraction of pole-like street furniture objects. The method simplifies the point cloud based on space voxelization. The horizontal sections are then analized to detect the circular shape pattern of the poles. González-Jorge et al. [12] exhibit an automated method for the detection of surface efflorescence in road overpasses. They detect the overpasses using geometrical considerations and extract information from the efflorescence using the radiometric information from the LiDAR. Holgado-Barco et al. [13] use mobile LiDAR data and computational algorithms to extract geometric parameters from the road (vertical profiles and cross-sections) to characterize the proper construction and road safety assessment.

Although there was a huge activity during last years for the automatic detection of road elements from mobile LiDAR data, unresolved situations still remain. One example is the automatic inventory of zebra crossing from mobile LiDAR data. Some successful attempts using image processing for the detection of zebra crossings have been done, although an accurate geo-positioning of the elements cannot be achieved [14], [15], [16]. Direct detection of zebra crossing from mobile LiDAR data with the corresponding geo-positioning information is not found in the bibliography. Zebra crossing is repetitive element in urban roads and their inventory and inspection are especially important for road safety. The manuscript is organized in the following form: Section 2 presents the study cases where the algorithm was validated and the LiDAR system used for data acquisition and the methodology used for automatic data management. Section 3 shows the results and discussion and finally the conclusions are presented in Section 4.