# Identification of Road Network Intersection Types from Vehicle Telemetry Data Using a Convolutional Neural Network

^{1}

^{2}

^{3}

^{4}

^{*}

## Abstract

**:**

## 1. Introduction

## 2. Literature Review

## 3. Data and Methods

#### 3.1. Data Collection and Preparation

**Collecting ground truth intersection locations:**From OSM’s road network, we obtained the location of several intersections in nine Canadian cities for which we observed a large number of trips. The intersections were manually labeled with the help of frequently enough updated images from Google Street View according to one of three categories: TL, ST or NS. The distribution of trips in each province was different depending on the number of UBI subscribers present. Moreover, as most TLs and NSs were positioned in arterial or collector roads, whereas STs were more present on local roads, more trips around TLs and NSs were available. Therefore, we balanced the dataset to have a total of 34 junctions of each type (each comprising two directions) for the network’s training. The breakdown of intersections by type and city is shown in Table 1, and the total number of trips in each type of intersections is given in Table 2.**Defining a junction window:**By exploring the speed variations of vehicles over many intersections, we chose to consider a window, as illustrated in Figure 1, of $d=50$ m before and after the center of the intersection. To set the value of d, we studied some trips individually in order to delimit an interval before and after each intersection, large enough to capture the driving pattern on various road topologies. The idea was to select a sufficiently large window to detect as much as possible any potential pattern in trip characteristics, but not too wide so as to contain points that might involve the characteristics of another intersection close to the one we wanted to classify. As reported by [26], the minimum intersection spacing between four-legged intersections along local roads or between adjacent intersections along collector roads is 60 m while on arterial roads, the minimum is 200 m. Therefore, choosing $d\le 50$ m prevented most points from the trips of a junction to appear in another. On the other hand, in order to capture the driving pattern, we needed to observe a sufficiently large number of trip points in the window. Note that the maximum speed over a city is around 50 km/h and recall that the average speed $\overline{v}$ may be computed as $\overline{v}=l/\Delta t$, where l is the driven distance and $\Delta t$ is the time interval. In our dataset, GPS observations were sampled each second, so we were able to estimate the minimum number of trip points to be about seven measurements when the window was $2d=100$ m wide, which was enough to detect driving patterns. Finally, for each intersection listed in Table 1 in each direction separately, trip points were filtered so that only trips that covered at least 80% of the junction window were kept.**Exploring trips to detect potential patterns in intersections:**First, we analyzed the distributions of speeds, accelerations and the duration of trips over the junction window by type of intersection (not shown). From this, it was seen that speeds and accelerations could potentially be useful to differentiate intersection types. In order to discover the possible differences between the three classes of intersections, the local estimates of the joint density of the speeds and accelerations are presented in the form of contour plots. To do this, we considered each heading direction on an intersection separately. As illustrated in Figure 2, these joint densities exhibit specific characteristic depending on the type of TCE at the junction. In the case of TLs (right-hand side of Figure 2), as expected, we observe roughly two modes, one at the speed zero and one around the maximum speed on the road. In the case of STs (left-hand side of Figure 2), nonzero accelerations are observed at low speeds between zero and 30 km/h (possibly reflecting slow transitions or incomplete stops of cars at stop signs). The middle plots in Figure 2 show that speed and accelerations at NS junctions are more similar to TLs than STs.Another way to visualize speed change along a given intersection is by plotting several individual driver trip speeds along the junction window. The resulting curves, examples of which are depicted in Figure 3, allowed us to detect similarities between speed profiles of trips at the same intersection type. Unlike NSs, speed curves at STs and TLs show a deflection towards zero as they approach the intersection center, however, the speed curves are closer to each other in the case of STs, creating a characteristic v shape visible in the left-hand side plot of Figure 3.**Preparing the dataset:**Each heading direction was considered separately. Only the trips that covered at least 80% of the bounding box formed by the junction window were selected. We discretized the trips in meters over an interval of length $2d=100$ m around the junction center ($d=50$ m was set in step 2). All paths then had the same dimension and could be represented as one hundred intervals of a distance of one meter each, in which we interpolated speed and acceleration values from the original data measurement. Examples for the speed are depicted in Figure 4.**Solving noise and errors found in our dataset:**Telematics trip data collected from dongles are usually more precise than those collected through smartphone applications, but they still are very noisy. Data cleansing is a necessary step to smooth the trajectories (i.e., using speed from GPS location stamp to interpolate positions when they are inaccurate or jumping), and to remove the invalid ones (i.e., missing points in the trips or erroneous speeds). Figure 5 shows the speed profiles of individual trips on a direction before (left) and after (right) this process.

#### 3.2. Proposed CNN Classifier for TCE at Junction

**Input:**Our network took as input D trips for an intersection in one direction, and two channels: speed and acceleration. We considered $D\in \{25,50,100\}$.**Convolution:**A $2d$ filter of size 3, 5 or 7 was applied to the inputs.**Batch normalization:**A $2d$ batch norm, as described in [27], helped to accelerate the training of the deep neural net.**Nonlinear:**Leaky-ReLU, introduced by [28], was the activation function and it is defined as$$Leaky.ReLU\left(x\right)=\left(\right)open="\{"\; close>\begin{array}{cc}0.01x,& x0\\ x,& x\ge 0.\end{array}$$**Pooling:**We downsampled the output signals by applying a $2d$ max-pooling.**Flattening and fully connected:**The last convolution layer was converted to a one-dimensional array, then passed to a fully connected layer. The output was three real numbers, which were the class scores corresponding to the three junction types.

#### 3.3. Other Methods for Classification of TCEs

## 4. Results

#### 4.1. CNN Model Training and Evaluation

#### 4.2. Metrics

#### 4.3. Functional Method with RF Training and Evaluation

#### 4.4. K2D Model Training and Evaluation

#### 4.5. Testing the Models

## 5. Discussion

- A four-way junction: (ST,ST) and (TL,TL), (ST,ST) and (NS,NS), or (ST,ST) and (ST,ST);
- A two-way junction: (ST,ST), (TL,TL) or (ST,NS).

## 6. Conclusions and Future Work

## Author Contributions

## Funding

## Data Availability Statement

## Acknowledgments

## Conflicts of Interest

## References

- Tselentis, D.; Yannis, G.; Vlahogianni, E. Innovative motor insurance schemes: A review of current practices and emerging challenges. Accid. Anal. Prev.
**2017**, 98, 139–148. [Google Scholar] [CrossRef] [PubMed] - Śliwiński, A.; Kuryłowicz, Ł. The Equilibrium on the Motor Insurance Market in Selected CEE Countries. In Economic Development and Financial Markets: Latest Research and Policy Insights from Central and Southeastern Europe; Śliwiński, A., Polychronidou, P., Karasavvoglou, A., Eds.; Springer International Publishing: Cham, Switzerland, 2020; pp. 143–164. [Google Scholar] [CrossRef]
- Stipancic, J.; Miranda-Moreno, L.; Saunier, N. Vehicle manoeuvers as surrogate safety measures: Extracting data from the GPS-enabled smartphones of regular drivers. Accid. Anal. Prev.
**2018**, 115, 160–169. [Google Scholar] [CrossRef] [PubMed] - Munoz-Organero, M.; Ruiz-Blaquez, R.; Sánchez-Fernández, L. Automatic detection of traffic lights, street crossings and urban roundabouts combining outlier detection and deep learning classification techniques based on GPS traces while driving. Comput. Environ. Urban Syst.
**2018**, 68, 1–8. [Google Scholar] [CrossRef] - Poudel, N.; Singleton, P.A. Bicycle safety at roundabouts: A systematic literature review. Transp. Rev.
**2021**, 41, 617–642. [Google Scholar] [CrossRef] - Leblanc, B.; Fouchal, H.; de Runz, C. Driver profile detection using points of interest neighbourhood. In Proceedings of the 2019 IEEE 90th Vehicular Technology Conference (VTC2019-Fall), Honolulu, HI, USA, 22–25 September 2019; pp. 1–4. [Google Scholar]
- Leblanc, B.; Ercan, S.; De Runz, C. C-ITS data completion to improve unsupervised driving profile detection. In Proceedings of the 2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring), Antwerp, Belgium, 25–28 May 2020; pp. 1–5. [Google Scholar]
- Raynal, L. Some elements for modelling updates in topographic databases. In Proceedings of the GIS LIS-International Conference, Denver, CO, USA, 19–21 November 1996; pp. 1223–1232. [Google Scholar]
- Carisi, R.; Giordano, E.; Pau, G.; Gerla, M. Enhancing in vehicle digital maps via GPS crowdsourcing. In Proceedings of the 2011 Eighth International Conference on Wireless On-Demand Network Systems and Services, Bardonecchia, Italy, 26–28 January 2011; pp. 27–34. [Google Scholar]
- Bruntrup, R.; Edelkamp, S.; Jabbar, S.; Scholz, B. Incremental map generation with GPS traces. In Proceedings of the 2005 IEEE Intelligent Transportation Systems, Vienna, Austria, 13–16 September 2005; pp. 574–579. [Google Scholar]
- Zourlidou, S.; Sester, M. Traffic Regulator Detection and Identification from Crowdsourced Data—A Systematic Literature Review. ISPRS Int. J. Geo-Inf.
**2019**, 8, 491. [Google Scholar] [CrossRef] - Golze, J.; Zourlidou, S.; Sester, M. Traffic Regulator Detection Using GPS Trajectories. KN J. Cartogr. Geogr. Inf.
**2020**, 70, 95–105. [Google Scholar] [CrossRef] - Hu, S.; Su, L.; Liu, H.; Wang, H.; Abdelzaher, T.F. Smartroad: Smartphone-based crowd sensing for traffic regulator detection and identification. ACM Trans. Sens. Netw. (TOSN)
**2015**, 11, 1–27. [Google Scholar] [CrossRef] - Wang, C.; Zourlidou, S.; Golze, J.; Sester, M. Trajectory analysis at intersections for traffic rule identification. Geo-Spat. Inf. Sci.
**2021**, 24, 75–84. [Google Scholar] [CrossRef] - Zourlidou, S.; Fischer, C.; Sester, M. Classification of street junctions according to traffic regulators. In Proceedings of the Geospatial Technologies for Local and Regional Development: Short Papers, Posters and Poster Abstracts of the 22nd AGILE Conference on Geographic Information Science, Limassol, Cyprus, 17–20 June 2019; Available online: https://agile-online.org/conference/proceedings/proceedings-2019 (accessed on 1 June 2022).
- Méneroux, Y.; Le Guilcher, A.; Saint Pierre, G.; Hamed, M.G.; Mustière, S.; Orfila, O. Traffic signal detection from in-vehicle GPS speed profiles using functional data analysis and machine learning. Int. J. Data Sci. Anal.
**2019**, 10, 101–119. [Google Scholar] [CrossRef] [Green Version] - Goyal, D.; Yuen, A.; Kim, H.S.; Murphy, J. Traffic Control Elements Inference using Telemetry Data and Convolutional Neural Networks. In Proceedings of the 8th International Workshop on Urban Computing (UrbComp 2019), SIGKDD Workshop, Anchorage, AK, USA, 5 August 2019. [Google Scholar]
- Müllner, D. Modern hierarchical, agglomerative clustering algorithms. arXiv
**2011**, arXiv:1109.2378. [Google Scholar] - Thavikulwat, P. Affinity propagation: A clustering algorithm for computer-assisted business simulations and experiential exercises. In Proceedings of the Conference of the Developments in Business Simulation and Experiential Learning—The Annual ABSEL Conference, Charleston, SC, USA, 5–7 March 2008; Volume 35, pp. 220–224. [Google Scholar]
- Yin, X.; Wu, G.; Wei, J.; Shen, Y.; Qi, H.; Yin, B. Deep learning on traffic prediction: Methods, analysis and future directions. IEEE Trans. Intell. Transp. Syst.
**2022**, 23, 4927–4943. [Google Scholar] [CrossRef] - Lv, Y.; Duan, Y.; Kang, W.; Li, Z.; Wang, F.Y. Traffic flow prediction with big data: A deep learning approach. IEEE Trans. Intell. Transp. Syst.
**2014**, 16, 865–873. [Google Scholar] [CrossRef] - Aqib, M.; Mehmood, R.; Alzahrani, A.; Katib, I.; Albeshri, A.; Altowaijri, S.M. Smarter traffic prediction using big data, in-memory computing, deep learning and GPUs. Sensors
**2019**, 19, 2206. [Google Scholar] [CrossRef] [PubMed] - OpenStreetMap Contributors. Planet Dump. Available online: https://planet.osm.org (accessed on 1 June 2022).
- Luxen, D.; Vetter, C. Real-time routing with OpenStreetMap data. In Proceedings of the 19th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, Chicago, IL, USA, 1–4 November 2011; ACM: New York, NY, USA, 2011; pp. 513–516. [Google Scholar] [CrossRef]
- Blais, P.; Badard, T.; Duchesne, T.; Côté, M.P. From Massive Trajectory Data to Traffic Modeling for Better Behavior Prediction in a Usage-Based Insurance Context. ISPRS Int. J. Geo-Inf.
**2020**, 9, 722. [Google Scholar] [CrossRef] - Transportation Association of Canada. June 2014 Errata to the Geometric Design Guide for Canadian Roads, 2014. Available online: https://tac-atc.ca/sites/tac-atc.ca/files/site/complete.pdf (accessed on 27 May 2022).
- Ioffe, S.; Szegedy, C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. arXiv
**2015**, arXiv:1502.03167. [Google Scholar] - Maas, A.L.; Hannun, A.Y.; Ng, A.Y. Rectifier nonlinearities improve neural network acoustic models. In Proceedings of the ICML, Atlanta, GA, USA, 16–21 June 2013; Volume 30, p. 3. [Google Scholar]
- Silverman, B.W. Density Estimation for Statistics and Data Analysis; CRC Press: New York, NY, USA, 1986; Volume 26. [Google Scholar]
- Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv
**2014**, arXiv:1409.1556. [Google Scholar]

**Figure 1.**Window of 100 m wide for a driver’s trip in one of the two heading directions of the road segment.

**Figure 2.**Speed–acceleration density contour plot for a given heading direction at two STs (

**left**), NSs (

**middle**) and TLs (

**right**).

**Figure 3.**Speed traces of trips over junction window at an ST (

**left**), an NS (

**middle**) and a TL (

**right**).

**Figure 4.**Speed measurements (solid dots) and interpolated values (empty dots) for one trip at an ST (

**left**) or at a TL (

**right**).

**Figure 5.**An example of speed profiles of trips on the window at an NS (

**top**) and an ST (

**bottom**), before (

**left**) and after (

**right**) filtering trips and fixing GPS errors.

**Figure 7.**Examples of kernel density images for the method of [17].

**Figure 8.**Training (blue) and validation (orange) losses per epoch for CNN_25 (

**left**), CNN_50 (

**middle**) and CNN_100 (

**right**).

**Figure 9.**Confusion matrix for class ST illustrating $T{P}_{ST}$, $F{P}_{ST}$, $T{N}_{ST}$ and $F{N}_{ST}$.

**Figure 11.**Training (blue) and validation (orange) losses per epoch for K2D_25 (

**left**), K2D_50 (

**middle**) and K2D_100 (

**right**).

Province | City | STs | NSs | TLs |
---|---|---|---|---|

Alberta | Calgary | 2 | 3 | |

Ontario | (Old) Ottawa | 5 | 2 | 3 |

Hamilton | 6 | 6 | ||

Toronto | 1 | 4 | 1 | |

Markham | 2 | |||

Québec | Lévis | 3 | 4 | |

Québec | 27 | 14 | 25 | |

Saint-Constant | 2 | 1 | ||

Total | 34 | 34 | 34 |

STs | NSs | TLs | |
---|---|---|---|

Number of trips | 157,291 | 401,532 | 924,528 |

**Table 3.**Twelve statistics of speed profiles calculated for each distance interval for the functional method.

F1 | Mean |

F2 | Standard deviation |

F3 | Third moment (coefficient of skewness) |

F4 | Fourth moment (coefficient of kurtosis) |

F5 | Sarle’s bimodality coefficient |

F6 | Median |

F7 | 15th percentile |

F8 | 85th percentile |

F9 | Dispersion (F8 − F7) |

F10 | Minimum |

F11 | Maximum |

F12 | Amplitude (F11 − F10) |

CNN | RF | RF_H | K2D | |
---|---|---|---|---|

25 trips | 97% | 96% | 95% | 91% |

50 trips | 98% | 97% | 96% | 95% |

100 trips | 99% | 98% | 97% | 96% |

D = 25 trips | D = 50 trips | D = 100 trips | |||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|

CNN | RF | RF_H | K2D | CNN | RF | RF_H | K2D | CNN | RF | RF_H | K2D | ||

ST | precision | 1.00 | 1.00 | 1.00 | 0.87 | 1.00 | 1.00 | 1.00 | 0.98 | 1.00 | 1.00 | 1.00 | 1.00 |

recall | 1.00 | 1.00 | 1.00 | 0.99 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 1.00 | 0.95 | |

F1-score | 1.00 | 1.00 | 1.00 | 0.93 | 1.00 | 1.00 | 1.00 | 0.99 | 1.00 | 1.00 | 1.00 | 0.97 | |

NS | precision | 0.93 | 1.00 | 1.00 | 0.91 | 0.97 | 1.00 | 1.00 | 0.88 | 0.97 | 1.00 | 1.00 | 0.94 |

recall | 0.98 | 0.88 | 0.86 | 1.00 | 0.98 | 0.92 | 0.89 | 1.00 | 1.00 | 0.94 | 0.92 | 0.97 | |

F1-score | 0.95 | 0.94 | 0.92 | 0.95 | 0.97 | 0.96 | 0.94 | 0.93 | 0.98 | 0.97 | 0.96 | 0.96 | |

TL | precision | 0.98 | 0.89 | 0.88 | 0.99 | 0.98 | 0.93 | 0.90 | 1.00 | 1.00 | 0.94 | 0.92 | 0.93 |

recall | 0.92 | 1.00 | 1.00 | 0.77 | 0.97 | 1.00 | 1.00 | 0.85 | 0.96 | 1.00 | 1.00 | 0.95 | |

F1-score | 0.95 | 0.94 | 0.93 | 0.87 | 0.97 | 0.96 | 0.95 | 0.92 | 0.98 | 0.97 | 0.96 | 0.94 | |

Overall accuracy | 0.97 | 0.96 | 0.95 | 0.91 | 0.98 | 0.97 | 0.96 | 0.95 | 0.99 | 0.98 | 0.97 | 0.96 |

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |

© 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Erramaline, A.; Badard, T.; Côté, M.-P.; Duchesne, T.; Mercier, O.
Identification of Road Network Intersection Types from Vehicle Telemetry Data Using a Convolutional Neural Network. *ISPRS Int. J. Geo-Inf.* **2022**, *11*, 475.
https://doi.org/10.3390/ijgi11090475

**AMA Style**

Erramaline A, Badard T, Côté M-P, Duchesne T, Mercier O.
Identification of Road Network Intersection Types from Vehicle Telemetry Data Using a Convolutional Neural Network. *ISPRS International Journal of Geo-Information*. 2022; 11(9):475.
https://doi.org/10.3390/ijgi11090475

**Chicago/Turabian Style**

Erramaline, Abdelmajid, Thierry Badard, Marie-Pier Côté, Thierry Duchesne, and Olivier Mercier.
2022. "Identification of Road Network Intersection Types from Vehicle Telemetry Data Using a Convolutional Neural Network" *ISPRS International Journal of Geo-Information* 11, no. 9: 475.
https://doi.org/10.3390/ijgi11090475