Automotive Innovation

, Volume 2, Issue 2, pp 110–120 | Cite as

Concept Study of a Self-localization System for Snow-covered Roads Using a Four-layer Laser Scanner

  • Tetsushi MimuroEmail author
  • Naoya Taniguchi
  • Hiroyuki Takanashi


Many advanced driver assistance systems have entered the market, and automated driving technologies have been developed. Many of them may not work in adverse weather conditions. A forward-looking camera, for example, is the most popular system used for lane detection but does not work for a snow-covered road. The present paper proposes a self-localization system for snowy roads when the roadsides are covered with snow. The system employs a four-layer laser scanner and onboard sensors and uses only pre-existing roadside snow poles provided for drivers in a snowy region without any other road infrastructure. Because the landscape greatly changes in a short time during a snowstorm and snow removal works, it is necessary to restrict the landmarks used for self-localization to tall objects, like snow poles. A system incorporating this technology will support a driver’s efforts to keep to a lane even in a heavy snowstorm.


Advanced driver assistance systems Adverse weather Laser scanner Self-localization system 

1 Introduction

Driving a car can be difficult in many regions that receive much snow in winter owing to poor visibility and road surface conditions. Extensive snow removal efforts are required to ensure road traffic can flow. When snowfall starts, visibility deteriorates and the landscape appears monotone, making it difficult for the driver to distinguish road boundaries. As the amount of snowfall increases, snow-removing vehicles push away snow from roadways, but in most cases the removed snow is simply shifted to shoulders of the road and covers roadside objects, such as traffic delineators, curbs, and guardrails. Heavy snowstorms sometimes lead to whiteouts in which the visibility falls to zero. Figure 1 shows an example of a suburban road in northern Honshu, Japan. The upper photograph shows curbs with snow poles on both sides and windbreak fences on the right side. The lower photograph shows the same road with poor visibility in a snowstorm, with snow piled to a height of 70 cm on the roadsides. Only snow poles and other tall objects can be seen.
Fig. 1

Photographs of a suburban road without snow (upper) and with snow (lower) in Akita prefecture, taken in 2016 and 2017

Self-localization systems of automated vehicles have been actively studied in recent years. Most of these systems comprise a high-definition three-dimensional map, a high-spec laser scanner, and a global positioning system (GPS) sensor [1]. The standardization of high-definition maps and the preparation of maps for an enormous number of routes are issues that need to be addressed urgently [2].

A classic map comprises purely static information and is updated once every few years. Meanwhile, high-definition maps require much more frequent updating because they are expected to include quasistatic information, such as information on road construction and winter closures, and dynamic information, such as information on nearby vehicles. However, it is still assumed that basic features of roads (e.g., roadside curbs, traffic signs, and guardrails) hardly change over a short period.

In snowy regions, snow accumulates in a short period on roadsides owing to snowstorms and snow removal works. Even a basic road feature will be instantly covered with snow unless it is tall, making it meaningless as high-definition map information. On roads in snowy regions, there is a strong need for not only automated driving but also driving support along a lane and snow removal assistance [3, 4].

The present paper proposes a self-localization system for snow-covered roads using a four-layer laser scanner, presents the results of a feasibility study, and proposes ideas for the development of an algorithm for the system.

2 Configuration of the Proposed System

The ultimate goal of the present research is to develop a driving support system that provides forward road boundaries to the driver even when roadside structures are covered with snow. Pre-existing roadside snow poles for drivers in a snowy region are employed without any other special infrastructure.

In recent years, a high-performance laser scanner with a 360° horizontal field of view (FOV) and many (e.g., 16 or 32) scanning layers has been used as the main onboard sensor in the development of automatic driving technology [5]. We employed a four-layer laser scanner in the present study. It is thought that a mid-performance sensor of this type can be widely applied as a sensor in an advanced driver assistance system owing to its potential to reduce costs.

Table 1 summarizes the main specifications of the four-layer laser scanner (SICK LD-MRS [6], equivalent to an IBEO LUX 2010) that was employed in our experimental vehicle. The left diagram in Fig. 2 illustrates the vertical FOV. The layers are referred to as the first layer, second layer, and so on in order from bottom to top. Hereafter, reflection points belonging to a layer will be indicated by the color shown in the figure. The right diagram illustrates the horizontal FOV. Arranging the sensor central axis along the longitudinal axis of the vehicle, the horizontal view angle is 35° to the left and 50° to the right. After installing the sensor on the vehicle, the vertical and horizontal viewing angles are adjusted using a reflection pole at a distance of 20 m in front of the vehicle. The range output is calibrated at the same time.
Table 1

Main specifications of the laser scanner [6]

Size (W × H × D)

164.5 mm × 88 mm × 93.2 mm

Laser class

Class 1

Laser wave length

905 nm


200 m/50 m (10% reflectivity)

Horizontal FOV

85° (+ 35°  to − 50°)

Vertical FOV

3.2°, 4 layers

Scan interval

40 ms

Fig. 2

Vertical (left) and horizontal (right) fields of view of the laser scanner [6]

The vehicle incorporates a GPS sensor, vehicle speed sensor, and inertial sensor as well as the laser scanner. The system uses a landmark map that is based on tall objects, like snow poles. The map is provided by road administrators or other road users who generate the map when driving through the section previously. Of course, the vehicle itself may generate the map in advance.

3 Layout of the Laser Scanner

The landscape expressed by laser reflection points greatly changes in a short time during a snowstorm and snow removal works, and it is thus necessary to restrict the landmarks used in self-localization to tall objects, such as snow poles. Hereafter, what we refer to as snow-pole-like objects are snow poles, traffic signs, street lamps, utility poles, and other objects that are taller than the surrounding piled snow and stand on the roadside with high reflectance.

Figure 3 shows that when the laser scanner is installed at a position lower than the height of the roadside piled snow, the snow pole cannot be detected because the foreground snow wall interrupts the laser beam. It is thus rational to mount the laser scanner higher than the predicted height of the snow wall.
Fig. 3

Laser sensor layout according to the height of snow piled at the roadside

Assuming a maximum snow height of 1.5 m and snow pole height of 2.5 m, installing a laser scanner on the roof of a passenger car (at a height of about 1.7 m) would reduce the occlusions of snow poles by snow walls and preceding vehicles.

The snow poles must be taller than the snow height determined by the maximum snowfall in a particular area. If the mounting height of the laser scanner corresponds to the high position in Fig. 3, landmark extraction restricted to snow-pole-like objects is straightforward in that the two reflections of a pole in the third and fourth layers have almost the same horizontal position.

4 Detection of Snow Poles in the Presence of Snowfall

Although millimeter-wave radar is robust against weather, its positioning accuracy (especially the resolution of lateral positioning) for a reflection is poor. Meanwhile, a camera does not function in poor visibility. A laser scanner has the advantages of high accuracy in measuring distance and a high angular resolution, but it does not necessarily perform well in poor weather because it uses near-infrared light. Rasshofer et al. designed a theoretical approach of using a laser sensor for various targets in rain, fog, and snow [7].

Laser scanners that are intended for outdoor use are sometimes incorporated with a multi-echo function such that they detect not only the first reflection of a laser beam but also a few successive reflections so that they might detect surrounding targets even during rain or snow. The laser scanner used in the present study detects up to three reflections.

The authors of the present study drew the following conclusions from the results of a previous study conducted under various weather conditions in winter [8]. There are many points of reflection from snowfall in the vicinity of the laser scanner during snowfall but far fewer points and lower reflection intensity at distances beyond 8 m (i.e., the (logarithmic) number of reflection points is inversely proportional to distance; see Fig. 4). Meanwhile, there is the possibility of detecting snow poles separated by more than 8 m because the reflection strength of a columnar roadside structure, such as a snow pole, is higher than snowfall and changes little with distance.
Fig. 4

Number of four-layer reflection points in the restricted area (0 m < x < 20 m, − 1 m < y < 1 m) for three levels of snowstorm [8]

Figure 5 shows an example of laser reflections recorded while driving on a road of a university campus at a speed of 30 km/h in a snowstorm of medium intensity. In the left figure, the vertical axis denotes the forward distance x and the horizontal axis denotes the right direction distance y, and the laser scanner is located at the origin. The reflection points are expressed by colored dots that correspond to the layer colors. It is observed that the reflections from snowfall are limited to the vicinity of the laser scanner while there are many roadside reflections at a distance. The right figure shows three snow poles extracted from the reflection points in the left figure; these poles are installed on the curb of the left road edge [8]. Snow pole reflections are extracted through the simple data processing described in Sect. 3.
Fig. 5

Upper: Forward image taken from the car traveling on a road, left: reflections from snowfall and roadside objects obtained using the laser scanner, right: snow pole reflections extracted from the left figure [8]

Figure 6 shows the accuracy of the interval between poles 1 and 2 on a campus road measured by the laser scanner on our experimental vehicle, which drove at a speed of up to 30 km/h, for four types of weather. The surveyed interval between anchors of snow poles P1 and P2 was 19.94 m. The average intervals slightly differ, probably owing to changes in inclination of the poles because measurements were made on different days except in the case of medium and heavy snowstorms. Despite variations among the averages, more than half of measurements are within 5 cm of their respective average value. The positioning accuracy of the detected pole is thus acceptable for each type of weather.
Fig. 6

Accuracy of the P1 − P2 interval measured by the laser scanner for four types of weather

Meanwhile, pole detectability deteriorates in poor weather. The left side of Fig. 7 shows the P1 detection rates for the same four types of weather considered in Fig. 6. The detection rate is defined as the number of scans in which a pole was detected divided by the total number of scans during a measuring period. The detection rate is lower in snowy weather and especially in a heavy snowstorm. This explains why the sample number N is only 2 in the case of the heavy snowstorm in Fig. 6, for which the simultaneous detection of P1 and P2 is required.
Fig. 7

Detection rate and recognition rate of the P1 pole when using the laser scanner for four types of weather

The right side of Fig. 7 shows the pole recognition rate after applying the tracking process that will be described in the third step of Sect. 5.2. The tracking process drastically improves the recognition rate in a heavy snowstorm.

5 Map Generation

5.1 Outline of Map Generation

A digital landmark map of snow-pole-like objects on the left road edge is generated using laser reflection data of tall roadside structures. The overall flow is presented in Fig. 8.
Fig. 8

Overall flow of map generation

In the upper left block, the four-layer laser scanner detects the relative positions of roadside snow-pole-like objects, which are taller than the expected maximum height of the snow wall, according to sensor coordinates. The road edge is regarded as being directly under the snow poles nearest the driving lane. These position data are converted into vehicle coordinates with the origin being at the center of gravity of the vehicle. The snow pole nearest the vehicle is referred to as the closest range pole (CRP, refer to Fig. 9), and the position of the CRP is passed to the coordinate transformation block.
Fig. 9

Poles and pole candidates

We here adopt digital odometry that uses the vehicle speed and yaw velocity obtained by onboard sensors. The vehicle motion and trajectory are calculated in real time in each driving section. The position of a CRP described in the vehicle coordinate system is transformed with the vehicle trajectory into ground-fixed coordinates having an origin at the section entrance of the road. We get a local map of the snow poles of the section. The local map is made part of a global map by adding the global position of the section entrance obtained by the GPS sensor.

5.2 Procedure for Detecting the Snow Pole Location

Figure 10 shows the detailed procedure of measuring the locations of snow poles in the upper left block of Fig. 8. The series of steps is performed within a scanning period of 40 ms.
Fig. 10

Procedure for measuring the snow pole location

  • Step 1 (uppermost step in Fig. 10): Eqs. (1), (2), (3), (4), and (5), given below, are used to estimate the current positions of previously stored poles and pole candidates. Here, Gx and Gy represent the position coordinates of a pole or a pole candidate in the vehicle coordinate frame with the origin at the center of gravity of the vehicle. Because the position is measured in sensor coordinates, an offset of the laser scanner from the vehicle center of gravity has been removed in advance. Gx and Gy on the right side of Eq. (1) refer to the previously measured position, while those on the left side refer to the estimated current position.

    Δx, Δy, and Δψ in Eq. (1) denote the movement of the vehicle during the scanning period Δt and are given by Eqs. (2), (3), and (4). Because the side slip angle β in Eq. (3) is not directly measured, it is approximated from the vehicle speed and yaw velocity according to the static relation expressed as Eq. (5) [9].

    $$\left[ {\begin{array}{*{20}c} {G_{x} } \\ {G_{y} } \\ \end{array} } \right] = \left[ {\begin{array}{*{20}c} {\cos \Delta \psi } & {\sin \Delta \psi } \\ { - \sin \Delta \psi } & {\cos \Delta \psi } \\ \end{array} } \right]\,\left[ {\begin{array}{*{20}c} {G_{x} - \Delta x} \\ {G_{y} - \Delta y} \\ \end{array} } \right]$$

    Δx, Δy: Longitudinal/lateral displacement during Δt.

    ΔΨ: Yaw angle inclement during Δt.

    $$\Delta x = x(\Delta t) = \int_{0}^{\Delta t} {V\cos (\beta + \dot{\psi }t){\text{d}}t} \cong \int_{0}^{\Delta t} {V{\text{d}}t = V\Delta t}$$
    $$\begin{aligned} \Delta y & = y(\Delta t) = \int_{0}^{\Delta t} {V\sin (\beta + \dot{\psi }t){\text{d}}t} \cong \int_{0}^{\Delta t} {V(\beta + \dot{\psi }t){\text{d}}t} \\ & = \int_{0}^{\Delta t} {V(f(V)\dot{\psi } + \dot{\psi }t){\text{d}}t} = V\dot{\psi }\left( {f(V)\Delta t + \frac{{\Delta t^{2} }}{2}} \right) \\ \end{aligned}$$
    $$\Delta \psi = \int_{0}^{\Delta t} {\dot{\psi }{\text{d}}t} = \dot{\psi }\Delta t$$
    $$\beta = f(V)\dot{\psi } = \left( {\frac{{b_{1} }}{V} - b_{2} V} \right)\dot{\psi }$$

    V: Vehicle speed.

    \(\dot{\psi }\): Yaw velocity.

    β: Side slip angle.

    b1, b2: specific constants for individual vehicles.

  • Step 2 The laser scanner executes a series of scans. Among the reflection points, each pair of third- and fourth-layer reflection points that are close to each other is set as a new pole candidate.

  • Step 3 This step is the so-called tracking process. Each new pole candidate detected in the second step is assessed in terms of whether it has a pole or pole candidate previously detected whose estimated current position is close to its actual position. If so, the probability of the candidate being a pole increases because the candidate can be considered as being re-detected. We call the candidate a “pole” if it is repeatedly detected. All poles and pole candidates are stored after the tracking process and reliability judgment are complete. The left side of Fig. 11 shows the laser reflections obtained on the same campus road as in Fig. 5 but without snow. White circles on the right denote the pole candidates obtained in the second step without being determined as poles in the third step, while black circles denote poles obtained in the third step.
    Fig. 11

    Extraction of pole candidates and poles

  • Step 4 The poles, whose detection reliabilities are confirmed in the tracking process, are classified into left-side poles and right-side poles of the roadway. This is an easy task because there is no pole on the roadway, which has a width of several meters (Fig. 12, left).
    Fig. 12

    Extraction of left-side poles (orange) and poles of left road edge (red)

To support travel in the left lane, only the poles along the left road edge are extracted from all left-side poles (Fig. 12, right). This restriction of using poles only of left road edge is unique, with other digital maps comprising many types of landmark. Executing this restriction requires certain techniques; however, the resulting poles of the left road edge are easy to use. Related discussions are presented in Sect. 5.3.

Step 5 When a pole of the left road edge comes within 10 m of the vehicle, it becomes a CRP, and the position of the CRP is passed to the coordinate transformation block. The CRP is transformed into ground-fixed coordinates to become map data.

5.3 Discussions on Poles of the Left Road Edge

To determine whether a pole is of the left road edge, it is verified that there is no left pole on the right side of the interpolation curve of poles of the left road edge. A circular arc with an appreciable radius is used as the interpolation curve. The success of the algorithm requires road edge poles to be spaced at appropriate intervals. In Fig. 13, for example, if the fourth pole is not installed, pole A is presumed to be a pole of the left road edge.
Fig. 13

Appropriate interval for poles of the left road edge

There is a possibility that buildings beside the road can also be used as landmarks. However, as well as buildings, natural or artificial slopes beside the road are commonly observed in snowy regions (see Fig. 14). These slopes and buildings are readily covered with snow, and laser reflections from them are expected to severely change. We therefore consider that it is better to restrict map landmarks to independently standing objects that are facing the roadway.
Fig. 14

Example of a snow-covered roadside (Fukushima prefecture, March 2018)

5.4 Map Generation

In the upper right block of Fig. 8, the CRP coordinates Gx and Gy are transformed from vehicle coordinates to ground-fixed coordinates using Eqs. (6) and (7). The vehicle yaw angle and trajectory in these equations are calculated using Eqs. (8), (9) and (10). A local map with an origin at the section entrance is obtained through this transformation. The global map is obtained by appending the GPS position of the section entrance to the local map.

The transformation of the CRP coordinates Gx and Gy from vehicle coordinates to ground-fixed coordinates is as follows.
$$\left[ {\begin{array}{*{20}c} {G_{x} } \\ {G_{y} } \\ \end{array} } \right] = \left[ {\begin{array}{*{20}c} {\cos \psi (t)} & { - \sin \psi (t)} \\ {\sin \psi (t)} & {\cos \psi (t)} \\ \end{array} } \right]\,\left[ {\begin{array}{*{20}c} {G_{x} } \\ {G_{y} } \\ \end{array} } \right]$$
$$\left[ {\begin{array}{*{20}c} {G_{x} } \\ {G_{y} } \\ \end{array} } \right] = \left[ {\begin{array}{*{20}c} {G_{x} + X(t)} \\ {G_{y} + Y(t)} \\ \end{array} } \right]$$
The calculations of the vehicle yaw angle and trajectory in ground-fixed coordinates are as follows:
$$X(0) = 0,\quad Y(0) = 0,\quad \psi (0) = 0$$
$$\psi (t) = \psi (t - \Delta t) + \Delta \psi$$
$$\left[ {\begin{array}{*{20}c} {X(t)} \\ {Y(t)} \\ \end{array} } \right] = \left[ {\begin{array}{*{20}c} {X(t - \Delta t)} \\ {Y(t - \Delta t)} \\ \end{array} } \right] + \left[ {\begin{array}{*{20}c} {\cos \psi (t)} & { - \sin \psi (t)} \\ {\sin \psi (t)} & {\cos \psi (t)} \\ \end{array} } \right]\,\left[ {\begin{array}{*{20}c} {\Delta x} \\ {\Delta v} \\ \end{array} } \right]$$
A local map of poles along the left edge of a campus road is illustrated in Fig. 15 with + symbols. The red circles show the surveyed positions of 21 poles. The map was generated from measurements made on a cloudy day, and all poles except P13 were detected. A countermeasure against such inevitable non-detection or misdetection, especially in poor weather, will be introduced in Sect. 6.3. The accuracy of pole positioning is generally good although small deviations are seen at the corner of the road.
Fig. 15

Generated map of poles of the left road edge along a campus road having a corner with a radius of 50 m

If the map is generated when there is no snow, reflection data of the first layer and second layer can be used to certify that the road edge poles are exactly on the road edge curb and to find obstacles on the road.

6 Self-localization

There have been many studies on the generation of a laser scanning point cloud through simultaneous localization and mapping based on LiDAR (light detection and ranging). The proposed system relies on the limited landmarks of roadside snow poles, because other roadside objects might be buried in snow or modified to have different shapes. This selection of a small number of data is the greatest point of difference from other approaches.

The GPS sensor plays a minor role in the proposed system. It is not fused with other sensors and is only used to select the map of the road section that is nearest the vehicle (see Sect. 6.1).

Too few poles are scanned in real time. To increase the number of available pole data, it is proposed to accumulate detected pole data (see Sect. 6.2) instead of expanding the laser field of view to the rear.

The limited number of map data for poles of the left road edge allows us to adopt a simple map matching algorithm (see Sect. 6.3), whereas the reliability and accuracy of map matching results become issues that will be discussed in Sect. 6.4. There is a standard for installing roadside delineators but not for installing snow poles. If the proposed system is to be widely applied, rules will be required for the installation interval, pole diameter, and pole height among other parameters. The system cannot identify the correct vehicle position in the case of a long straight road with snow poles placed at uniform intervals. This means that a rule for setting snow poles at intentionally irregular intervals should be included.

6.1 Map Selection

The vehicle that will receive the lane guidance service shall have map data comprising the positions of poles of the left road edge for each road section according to the traveling direction. When the vehicle approaches a certain road section, the system reads the map data, triggered by the condition that the current GPS position approaches the global position of the section entrance recorded on the map.

6.2 Accumulation of Position Data for the Rear Pole

The installation interval of snow poles on arterial roads is around 10–30 m even for relatively dense distributions, and it is expected that at most three poles of left road edge will be detected at a time by the laser scanner. Because it is difficult to execute highly accurate self-localization using only position information of a few poles in real-time detection, the position information of poles that have passed behind the vehicle is also used. Instead of detecting the rear poles, each CRP position is accumulated through coordinate transformation using Eq. (1). Figure 16 presents the situation. The matching success rate is expected to improve as the number of accumulated poles increases.
Fig. 16

Accumulation of data of rear pole positions

6.3 Map Matching

Map matching is performed between the pole position information Pi of the map (in ground-fixed coordinates) and the pole position information Qj (in vehicle coordinates) that the system has accumulated or detects in real time. The system searches for the best matching pole row to perform self-localization estimation (Fig. 17).
Fig. 17

Map matching for the pole position that the system has accumulated or detects in real time

An extreme deterioration of matching due to the non-detection of a pole (exemplified by the interval between Q3 and Q4 in Fig. 17) or misdetection is overcome through so-called elastic matching (j = ui), which finds the optimal correspondence between the two series. ui is referred to as the warping function.

We apply the well-known technique of dynamic programming matching (DPM) as follows. DPM is commonly used in various fields, such as voice recognition, character recognition, and video spotting [10].

Letting di(j) be the cost function of associating pole Pi with pole Qj, the optimization problem is formulated as follows, where F is the matching cost depending on ui and iSTART.
$$\begin{array}{*{20}l} {\text{minimize}} \hfill & {F = \sum\limits_{{{i = i}_{\text{START}}}}^{{{i}_ {{\text{MAX}}^{ - 1}} }} {d_{i} (u_{i} )} } \hfill \\ {{\text{w}} . {\text{r}} . {\text{t}}} \hfill & {u_{1} ,u_{2} , \ldots } \hfill \\ \end{array}$$
subject to the routes described in Fig. 18.
Fig. 18

Trellis for DP matching

The numbers i and j are used as sampling times in many DPM applications, whereas they are pole numbers in our case. The calculation cost must therefore be low.

When the Pi mapping is perfect, the three inclination values of 2, 1, and 1/2 in Fig. 18, respectively, correspond to Qj detection statuses of misdetection (or false positive in detection), successful detection, and non-detection (or false negative in detection). The three inclinations are, respectively, denoted by D, S, and H.

The cost function di(ui) is given by Eq. (12) when the inclination is equal to 1. Here, vector pi is defined with the direction of pi−1 to have a local feature quantity as shown in Fig. 17. qj is defined in the same manner.
$$d_{i} (u_{i} ) = \left\| {\varvec{p}_{i} - \varvec{q}_{{u_{i} }} } \right\|^{2}$$

The DP matching calculation becomes more expensive as the given iMAX and jMAX increase. In the example below, iMAX is set within 21 by extracting the neighboring pole row from the current map, and jMAX is set within 8. When jMAX is 8, the route whose inclinations are all 1 is expressed as SSSSSSS with seven S symbols that correspond to a series of seven black arrows from (iSTART, 0) in Fig. 18.

A DP matching result is shown in Fig. 19. In this example, the surveyed data of 21 poles in Fig. 15 are employed as Pi and the generated data of eight poles from the counterpart of P10 in the figure as Qj. The lateral axis denotes iSTART. The vertical axis denotes the matching cost. Eight curves are shown corresponding to the eight routes of HSSSSSS, SHSSSSS, SSHSSSS, SSSHSSS, SSSSHSS, SSSSSHS, SSSSSSH, and SSSSSSS selected from 308 routes in the trellis of Fig. 18. The minimum cost is attained at iSTART = 10 for SSHSSSS, which corresponds to the Qj whose counterpart of P13 cannot be detected. It is thus shown that correct matching can be achieved even when the measurement includes a non-detection. The algorithm based on the trellis in Fig. 18 can be applied to data with up to three non-detections and/or misdetections.
Fig. 19

DP matching result

6.4 Discussions on the Accuracy and Reliability of Self-localization

Because the CRP plays the main role in the system, the positioning accuracy and detectability of poles close to the laser scanner (within 10 m or so) are the most important. Figure 6 in Sect. 4 shows that the positioning accuracy of the detected pole is good, but the pole detection rate deteriorates in poor weather. The tracking technology described in Sect. 5.2 can improve the pole recognition rate, but there remain considerable non-detection possibilities. It is expected that excessive enhancing in the tracking process so as not to lose past detected poles will increase misdetection possibilities. DPM with elastic matching described in Sect. 6.3, which allows for false detections, is thus required.

In terms of ensuring the system can be used on various types of actual road and in various types of weather, it is expected that many improvements will be made to all technologies of the system, especially the tracking process and matching algorithm. The effectiveness of these technologies should be verified under such varying conditions.

7 Conclusions

We proposed a self-localization system that is useful even when roadside structures are covered with snow and presented results of a feasibility study and ideas for developing an algorithm of the system.

As a main sensor, we used a four-layer laser scanner that can be widely adopted because of its low cost and that can detect the positions of snow-pole-like objects taller than the predicted height of a snow wall. To support travel along the left lane of a road, among the snow-pole-like objects, only poles near the left road edge are extracted. Because the map generated in advance also consists of poles only on the left road edge as landmarks, it is possible to estimate the forward road boundaries covered with snow by matching the position information of the detected poles with a map.

We examined four items in confirming the feasibility of the proposed system.
  1. (1)

    A laser scanner layout for the detection of snow poles in piled snow.

  2. (2)

    The detectability of a snow pole in snowfall.

  3. (3)

    Generation of a map with snow poles as landmarks.

  4. (4)

    Self-localization estimation through map matching.


Section 3 showed that, by installing a four-layer laser scanner on the vehicle roof, it was possible to reduce the occurrence of occlusion and to detect pole candidates with a simple algorithm using only the reflection-point information of the third and fourth layers of the scanner.

Section 4 drew the following conclusions from the results of experiments conducted in various weather conditions during winter. There are many points of reflection from snowfall near the laser scanner during snowfall but far fewer and lower reflection intensity beyond 8 m. Meanwhile, because the reflection strength of a roadside structure such as a snow pole is high and does not change greatly with distance, it is possible to detect the positions of poles more than 8 m apart. The pole position accuracy is acceptable in four types of weather, although pole detectability deteriorates in poor weather. This low detectability in poor weather is improved by the proposed tracking process.

Section 5 reported the results of a detailed investigation of the tracking algorithm for the stable detection of snow poles and extraction algorithm of poles near the left road edge. It was shown that it is possible to develop a map generation function.

Section 6 showed that self-localization estimation can be realized by matching a previously generated map and pole position information recorded in real time. The introduced elastic matching provides correct matching even when the data include false detections.

The information necessary for a driver to keep in a lane is the road alignment ahead of the vehicle. It is possible to provide the driver with the alignment of the road forward without being limited by the detection range of the laser scanner because the map retains information of the road edge position beyond the detection range of the laser scanner. A lane guidance function can therefore be constructed by expressing the edge position of the road forward relative to the vehicle on the onboard display using the vehicle position estimation system.

8 Future Work

The experimental vehicle shown in Fig. 20 is being modified to demonstrate the operation of the proposed system. In this system, a four-layer laser scanner is installed at the front end of the roof (at a height of 1.79 m) (Fig. 20, left).
Fig. 20

Appearance of the experimental vehicle

In addition, as an application of the above research, there are plans to develop a cross-sectional measurement system of a snowy road for the purpose of supporting snow removal work. A similar prior study using a laser scanner confirmed that the effective road width can be measured with sufficient practical accuracy [11]. We plan to develop a system that combines the cross-sectional measurement function using a dedicated two-dimensional laser scanner installed at the rear end of the roof (Fig. 20, right) with the support system for driving in the left lane presented in this report.



We would like to express our sincere gratitude for Mr. Kenji Kimura’s kind advices and cooperation in measuring and data processing. This work was supported by Japan Institute of Country-ology and Engineering (JICE 2017 and 2018).

Compliance with Ethical Standards

Conflict of interest

On behalf of all the authors, the corresponding authors state that there is no conflict of interest.


  1. 1.
    Kato, M., et al.: Development of a self-localization technology with LiDAR Data processing utilizing landmark information. Trans. JSAE 49(1), 112–117 (2018)Google Scholar
  2. 2.
    Dynamic map construction study consortium, “SIP-adus study of prototyping and evaluation for dynamic map construction to realize automated driving system”, Entrusted research report in Japanese, Mar (2017). Accessed 31 Jan 2018
  3. 3.
    Watanabe, S., Hayashi, M., Murakami, K.: Measures to raise the sophistication of expressway snow and ice management. In: 17th ITS world congress, TP-23-1 (2010)Google Scholar
  4. 4.
    Norem, H., Thordarson, S.: Winter maintenance practice in the northern periphery, report of ROADEX sub project B phase I (2001)Google Scholar
  5. 5.
    Velodyne LiDAR Homepage. Accessed 19 Sep 2018
  6. 6.
    SICK AG: LD-MRS laser scanner operating instructions (2009)Google Scholar
  7. 7.
    Rasshofer, R.H., et al.: Influences of weather phenomena on automotive laser radar systems. Adv. Radio Sci. 9, 49–60 (2011)CrossRefGoogle Scholar
  8. 8.
    Taniguchi, N., Mimuro, T.: Detection of snow pole on snow-covered road using four-layer laser scanner. Trans. JSAE 49(2), 341–346 (2018)Google Scholar
  9. 9.
    Mimuro, T., Basic theory of vehicle dynamics. In: Iguchi, M. et al. (eds.) Automotive technology handbook, vol. 1, pp. 287–297. JSAE, Tokyo (2015)Google Scholar
  10. 10.
    Uchida, S.: DP matching—fundamentals and applications -, technical report of IEICE, PRMU2006-166 (2006)Google Scholar
  11. 11.
    Sato, S., Sumita, N., Murakami, K.: An efficient measurement technique for winter road width. In: 14th ITS Japan symposium (2016)Google Scholar

Copyright information

© China Society of Automotive Engineers (China SAE) 2019

Authors and Affiliations

  1. 1.Faculty of Systems Science and TechnologyAkita Prefectural UniversityYurihonjoJapan
  2. 2.Graduate School of Systems Science and TechnologyAkita Prefectural UniversityYurihonjoJapan
  3. 3.College of EngineeringNihon UniversityKoriyamaJapan

Personalised recommendations