1 Introduction

In the winter time, slippery road conditions, such as snowy, icy or slushy pavements, become a latent danger for road safety. More than 70% of the roads in the United States are located in snowy areas, with an average annual snowfall of over 5 inches. In addition, nearly 70% of the American population resides in regions with snows. Ice and snow would diminish pavements' friction and vehicles' maneuverability, thus resulting in reduced vehicle speeds, lowered roadway capacities, and increased accident risks. Under the effect of poor visibility and snowy or icy pavements, average arterial speeds are reduced by 30 to 40%, while freeway speeds are reduced by 5 to 40% (Pisano et al., 2008). Meanwhile, accident rates rise dramatically. For example, 24% of weather-related vehicle crashes occur on snowy or icy pavements, causing more than 1,300 deaths and 116,800 injuries each year. To address these problems, approximately 20% of the maintenance budget of US Department of Transportation (DOT) is spent on winter road maintenance. Therefore, mapping the slippery pavement conditions in real time will help deploy proper countermeasures, such as snowplowing, spreading of deicing chemicals and shutting down roads, so as to avoid traffic jams and accidents. More importantly, information about slippery roadway conditions can be used to alert drivers to operate cautiously and avoid accidents when running on slippery pavements. However, the conventional methods to survey pavements' slippery conditions (such as measuring road frictions by using locked wheels with threadless tires) are challenging, as they would not only require significant manpower and resources, but also can be dangerous for the operation of special pavement survey equipment.

As an emerging technology, the connected vehicle (CV) technology is potentially transforming the transportation industry. This technology is a combination of hardware, software and firmware, and it allows for the dynamic transmission of messages between vehicles (V2V) and between vehicles and infrastructure (V2I) (Lu et al., 2014). The types and components of CV messages are specified in the Society of Automotive Engineers (SAE) J2735 standard and are broadcast via dedicated short range communication (DSRC) or potential 5G technology (Standard, 2009). CV's safety applications are built around the Basic Safety Messages (BSM), which consists of two parts: Part 1 contains the core data elements, including: size, position, speed, heading, acceleration, steering wheel angle and other information of vehicles; and Part 2 consists of a number of optional components, such as air temperature, wiper and lamp status, and activation of traction control system (TCS) and anti-lock braking system (ABS). However, not all of the parameters in Part 2 are currently available from vehicles, and some of them are not authorized by USDOT for sharing. Currently, American Association of State Highway and Transportation Officials (AASHTO) has established an infrastructure blueprint to achieve a mature connected vehicle environment by 2040. By then, most of the vehicles on roads will be connected, with up to 80% of traffic signals and 25,000 other roadside locations being V2I-enabled. Furthermore, accurate, real-time, and localized traveler information will also be available on 90% or more of roadways (Wright et al., 2014). Generally, CV can transmit data at a frequency of 10 Hz via V2V and V2I; therefore, the widespread deployment of these communication technologies would provide an unprecedented amount of data, even at a low CV market-penetration rate. Such “big data,” if combined with the emerging deep learning algorithms, would be well suitable for detecting pavement-related hazards imposed by winter weathers.

This study aims to exploit the potential of using CV data to map pavements' slippery conditions in real time, while providing guidance for proactive countermeasures, so as to prevent potential crashes at intersections. The schematic of the detection system proposed in this study is shown in Fig. 1. By deploying roadside equipment (RSE) at an intersection, data from the surrounding CVs can be collected and processed by a pre-trained deep learning model, in order to predict the slippery level of pavements at the intersection. The predicted slippery conditions of pavements can then be broadcast (by RSE via DSRC or 5G technology) to vehicles (either human-operated or autonomously driving), so as to promote cautious driving behaviors.

Fig. 1
figure 1

Schematic of the real-time slippery pavement detection and warning system based on CV and RSE at an intersection (only CV and RSE are shown for illustration purpose. Regular vehicles, which account for a large portion of the traffic, are not shown)

Full size image

The rest of this paper is organized as follows: In Sect. 2, the currently published studies focused on slippery detections and their corresponding limitations are presented; in Sect. 3, details about traffic simulation and CV data acquisition with VISSIM are described; in Sects. 4 and 5, the classification and warning models proposed in this study are introduced; in Sect. 6, the models' performance and safety improvement from the simulation results are discussed; and finally, in Sect. 7, this study is concluded.

2 Literature review

Over the past years, many studies have been undertaken to monitor pavement conditions under winter weathers. Federal Highway Administration (FHWA) has developed a Road Weather Information System (RWIS) (Andrey et al., 2001), which plays a vital role in monitoring road weather conditions. It consists of a central system, a communication system and numerous Environmental Sensor Stations (ESSs). The ESSs are deployed in the field for data measurements about weathers, pavement conditions, etc. The collected data are then transmitted to the central system via the communication system finally develop nowcasts and forecasts. The processed RWIS data are useful for road operation and pavement maintenance. Similarly, Linton and Fu (2016) have proposed a Road Surface Condition Monitoring System (RSCMS), which combines pavement images with data from road weather information centers to monitor pavement conditions in real time. By investigating three machine learning classification methods, i.e., Artificial Neural Network (ANN), Classification and Regression Trees (CART) and Random Forest (RF), for potential applications, they evaluated these methods' performance in field tests, finding that all models achieved proper classification rates; especially, RF delivered the best classification performance. In addition, many other existing researches have also used vehicle-related data for detection of slippery road conditions (Enriquez et al., 2012; Heimann et al., 2006; Hou et al., 2017; Irschik & Stork, 2014; Padarthy & Heyns, 2019; Panahandeh et al., 2017).

Panahandeh et al., (2017) have demonstrated how to use vehicle data and weather report data to predict road friction levels. They formulated the prediction problem as a binary classification (slippery or non-slippery) task and introduced three machine-learning based methods, i.e., support vector machine (SVM), ANN and logistic regression. The experimental results demonstrated that their proposed methodology could successfully recognize whether or not pavements are slippery. Irschik and Stork (2014) have designed and implemented a pavement hazard identification system, which is able to classify the then-current pavements based on standard vehicle sensor data and send a warning to drivers, so that they can react in time when approaching a dangerous roadway. The experimental results demonstrated an excellent detection rate. Enriquez et al. (2012) have developed a programmable and expandable On-Board Diagnostics (OBD) device, which is capable of not only reading various vehicle data, but also detecting whether or not any vehicle slippages have occurred in real time. Hou et al. (2017) have proposed a framework to detect and report slippery road conditions by using smartphones and OBD-II adapters. When vehicles skid, a mismatch would be found between the wheel speed and the ground speed (which is the vehicle's actual running speed); this information can be used as a detector to monitor vehicles' skidding events. A skidding detection algorithm was developed and applied in field tests conducted at Buffalo, New York, and the evaluation results showed that the algorithm achieved satisfactorily high accuracy with a very low false-positive rate. Heimann et al. (2006) and Padarthy and Heyns (2019) have also presented similar systems for slippery recognition and friction measurement.

However, the above studies or proposed systems all have certain limitations. For example, although RWIS and RSCMS have been put into operation on real roads, they only reliably cover some limited areas adjacent to where they are installed, leading to fragmented information about the roads that are affected by inclement weather conditions. Most vehicle-based detection systems rely heavily on ABS or TCS data from vehicles, but such data are considered proprietary by vehicle manufacturers and are not publically available currently. Additionally, although many studies have deployed vehicle fleets to collect data and detect road surface conditions, the cost of their fleets is much high and the fleets are incapable of covering the whole transportation network. In order to solve these issues, a CV-based slippery pavement detection system is proposed in this study. In this system, the data of common vehicle messages transmitted by connected vehicles are used to provide real-time and complete coverage of pavements' slippery conditions in the whole road network. And the system also utilizes the data from Part 1 of BSM, which is available and can be easily deployed for field applications. As one of the most popular deep learning models for time series of data, long short-term memory (LSTM) is used to process data and predict pavements' slippery conditions. The process of preprocessing CV messages and detecting slippery conditions can be automated. Real-time alerts to drivers about pavements' slippery conditions can trigger their actions to prevent accidents. This system can also be used to make road management and support winter maintenance decisions.

3 Traffic simulation with VISSIM

This study uses VISSIM for traffic simulation, system implementation and evaluation. VISSIM is adopted due to its advantages in reproducing realistic traffic networks and vehicle dynamics. In addition, it allows considering the influence of various factors that may affect the traffic network conditions, including: drivers' behaviors, inter-vehicle interactions, environmental conditions, network characteristics, etc. Holistic simulations with such tools as VISSIM can overcome the limitations of field-based studies, which are typically constrained with insufficient data acquired, high implementation cost, and uncontrollable nature of the climate. VISSIM has been proven to be able to achieve satisfactory performance based on validations in several field studies (Asamer et al., 2013; Chen et al., 2019b). Moreover, it provides a simulation platform for CVs and a COM interface for users to program and embed algorithms. With VISSIM, pavement conditions, impacts of weather conditions, CV BSM data, and human driving behaviors can all be simulated. For these considerations, the real-time slippery detection system proposed in this study would be implemented and evaluated on the VISSIM simulation platform.

3.1 Basic road map design

For the traffic simulation, the base road map is drawn in reference to the Ann Arbor Safety Pilot Model Deployment (SPMD) program, which was conducted in Ann Arbor, Michigan by USDOT from August 18, 2011 to August 29, 2014. The field test includes 75 miles of instrumented roadways. Approximately 28 sets of roadside equipment (RSE) capable of receiving vehicle data and broadcasting road information and 3,000 equipped vehicles capable of V2I communication via DSRC are deployed throughout the network. Most of the RSE devices are placed at signalized intersections, while the remaining ones are installed at curves and freeway locations. In this study, only a portion of the researched field is reproduced in VISSIM, as shown in the left roadmap of Fig. 2.

Fig. 2
figure 2

(Left) Google map of the study area of the Ann Arbor SPMD program; (right) Road map created in VISSIM based on a portion of the SPMD program (the circles indicate the locations of roadside units and the ranges of communications)

Full size image

The simulated area is circled in red color on the Google map, which includes five roadways: Plymouth Rd, Maiden Ln, Fuller Rd, Glazier Way, and Huron Pkwy. The roadmap in the right side of Fig. 2 indicates the roadways created with VISSIM. It is assumed that there are totally six RSE devices deployed at each signalized intersection, as shown in the six circles at the crossings. The radius of each circle is set to 100 m, representing the range of V2I communication (Standard, 2009).

3.2 CV tire-road interaction model

Vehicles can easily skid on snowy and icy pavements when starting or braking. The existing research has shown that the acceleration, ground speed (i.e., the vehicle's driving velocity relative to the ground), and wheel speed (i.e., the vehicle's velocity inferred from the rotations of the wheel) are important parameters in detecting pavement conditions (ASTM, 2011; Heimann et al., 2006; PROKEŠ, 2015; Wang et al., 2004; Zhao et al., 2017). When skidding, the vehicle's wheel speed is significantly different from the ground speed, as shown in Fig. 3 (Hou et al., 2017). The slippery and non-slippery conditions can be described in a tire-road interaction model for the kinetics and kinematics of vehicle road interactions, as illustrated in Fig. 4 and Eqs. (1)–(4).

$$\mu = \frac{{F_{{\text{x}}} }}{{F_{{\text{z}}} }}$$
(1)
$${\text{ma}} = F_{{\text{x}}} - R_{{\text{x}}} - F_{{\text{a}}}$$
(2)
$$I_{{\text{w}}} \dot{w} = T_{{\text{x}}} - r_{{\text{w}}} F_{{\text{x}}}$$
(3)
$$V_{{\text{w}}} = r_{{\text{w}}} w$$
(4)

where, a indicates the acceleration; µ indicates the tire-friction coefficient; Fx and Fz indicate the longitudinal and normal forces acting on all tires, respectively; m indicates the total vehicle mass; Rx indicates the rolling resistance; Fa indicates the aerodynamic drag; Iw indicates the wheel inertia; w indicates the angular velocity; ẇ indicates the angular acceleration; Tx indicates the drive or brake torque delivered to the wheel; Vw indicates the wheel velocity; Vg indicates the ground velocity; and rw indicates the tire radius.

Fig. 3
figure 3

Speed comparison for a typical vehicle skidding event. a Skidding during acceleration; b Skidding during deceleration

Full size image
Fig.4
figure 4

Schematic of vehicle's longitudinal dynamics

Full size image

The rolling resistance and aerodynamic drag vary with vehicles, weathers and pavement types, and they are usually much smaller than the longitudinal and normal forces. Thus, these two parameters can be discarded. Therefore, Eqs. (1)–(2) can be simplified as:

$$a = \mu g$$
(5)

Equation (5) indicates that the tire-pavement friction limits the maximum vehicle acceleration. This explains why the maneuverability of vehicles is compromised when vehicles travel on icy or snowy pavements. This tire-road interaction model is used in this study to simulate vehicle behaviors under different pavement conditions. Since the effects of forces in uphill or downhill are not considered, this model is applicable for vehicles traveling along a straight line on a leveled road. Therefore, this study only focuses on evaluating the performance of the CV system framework and algorithms in detecting slippery pavements along straight and flat roads. The detection process on curved roads or along slopes is more complex and will be addressed in future researches.

3.3 Simulation under different pavement conditions

Based on the pavement friction coefficient, pavements can be divided into three categories: dry, snowy and icy, with the dry surface corresponding to the least slippery condition and the icy surface to the most slippery condition. The slippery detection process can then be formulated as a classification problem for deep learning purpose. According to the statistical data of road survey results (Ahn et al., 2011; Lerner et al., 2015; Olney & Manorotkul, 2007), the ranges of friction coefficient, the impact on driving and the preventive measures for corresponding pavement conditions are summarized in Table 1.

Table 1 Description of different pavement surface conditions
Full size table

In the simulation, the friction coefficients of each roadway are randomly selected from the ranges listed in Table 1, and they are assumed to change every hour, so as to ensure the generality of the proposed system. The traffic volumes on each simulated road and the traffic flow distributions in every merge or diverge section are inputted based on the historic traffic data for January of 2013 retrieved from Michigan DOT traffic information system. The annual average daily traffic (AADT) of Plymouth Rd, Maiden Ln, Fuller Rd, Glazier Way and Huron Pkwy are 24,750, 3095, 19,610, 8,845 and 5,005, respectively. The speed limit is set to 60 km/h for all roadways. The average speeds (km/h) of vehicles on dry, snowy and icy pavements are empirically set to 60, 42 and 36, respectively. Both CVs and normal vehicles are simulated, but only CVs are capable of transmitting data to RSE devices at 10 Hz sampling rate and receiving warnings from RSE devices. Once snowy or icy pavements are detected, the corresponding warnings are assumed to be sent by RSE devices to the surrounding CVs via DSRC at 1 Hz update rate. It is also assumed that drivers of the CVs receiving the warnings will change the vehicle operation by taking the precautions as described in Table 1 (i.e., to reduce vehicles' speeds and increase vehicles' following distance). Given that this study is focused on the system development, the number of CVs is set to be the same as the number of normal vehicles (i.e., the CV market penetration rate is 50%). The performance of the system at different CV market penetration rates will be discussed in future work.

4 The proposed classification model

A deep-learning based model is developed to classify pavement conditions based on the simulated CV data. Deep learning has been implemented successfully in many domains, such as computer vision, natural language processing, speech recognition and vehicular network (Chen et al., 2018a, 2018b, 2019a; Goodfellow et al., 2016; Graves et al., 2013; Hinton & Salakhutdinov, 2006; Hinton et al., 2012; LeCun et al., 2015; Liang et al., 2018). In this study, long short-term memory (LSTM), a kind of artificial recurrent neural network (RNN) architecture, is used to analyze the time series characteristics of vehicle data. In RNN, connections between nodes form a directed graph along a temporal sequence, so as to exhibit temporal dynamic behaviors. Unlike feedforward neural networks, RNN can use internal states to process input sequences. This feature makes RNN suitable for tasks with sequences of data, such as speech recognition, language translation and image captioning (Chen et al., 2018a, 2018b; Hochreiter & Schmidhuber, 1997; Sak et al., 2014; Zhang et al., 2019). LSTM is one of the most successful RNN networks, because it is able to deal with the exploding and vanishing gradient problems generated by long sequence datasets.

4.1 The framework of the model

The constructed network architecture of LSTM is shown in Fig. 5. Each LSTM cell consists of an input gate, a forget gate, and an output gate. The input gate controls the extent a new value flows into the cell, the forget gate controls the extent a value remains in the cell, and the output gate controls the extent a value in the cell is used to compute the output activation of the LSTM cell. The operation of each gate is determined by weights, which are learned during training. Different LSTM cells share the same set of weights.

Fig. 5
figure 5

The information flow of LSTM network used in this study

Full size image

4.2 Training of the model

The data used for the LSTM model's training is generated from VISSIM simulations. For each pavement slippery condition scenario, the simulations would run for 6 h. In the dry condition scenario, all the routes were simulated. However, in the snowy or icing condition scenario, only the places where vehicles are likely to skid, such as intersections and curves, were simulated. The CV data from the BSM utilized in the training of the LSTM model mainly include: vehicle velocity, acceleration, heading, longitude, latitude, etc. The vehicle kinematics data include: vehicle ground speed, wheel speed, ground acceleration, and wheel acceleration, all of which are used for training the slippery detection model, while vehicle heading, longitude and latitude are used for localization.

Firstly, a standardization process is applied to all simulated vehicle kinematics data by bringing the data into a common format with a mean of 0 and a standard deviation of 1. Then, data segmentation is conducted to divide the whole dataset into multiple windows of t-second duration. For a 10 Hz sampling rate used by CVs, each window would contain 10t data points. The 10t is the length of the time sequences used for slippery detection. For the LSTM model, 10t is the number of LSTM cells. Given the fact that the longer the sequences used for detection, the richer the features involves, but the worse the timeliness of detection, t is empirically set to 1 s (Hou et al., 2017). It is observed that segmentation at a 1-s interval would contain sufficient information for capturing the CV skidding process and delivering timely warnings and reaction time to the next CV. Afterward, each sequence is labeled as dry, snowy or icing class according to the corresponding pavement condition. Finally, combined with corresponding labels, the CV data sequences are used for model training. Table 2 shows the number of sequences under each of the three different pavement conditions. It can be noticed that the dataset contains obvious imbalance for different pavement conditions, which may result in a prediction bias towards the class with more data (i.e., the class of dry pavements). Therefore, the under-sampling and over-sampling techniques are applied to the raw dataset to generate a new balanced sequential dataset, in which the number of sequences of each class is 282,197. The balanced dataset is then split into a training set and a validation set, with 80% for training and 20% for validation. The stratified sampling method is used to ensure that the training and validation sets have approximately the same percentage of samples of each target class in the complete set. The training set is used for LSTM training, while the validation set is used for hyperparameters optimization.

Table 2 Number of 1-s vehicle kinematics sequences for each type of pavement conditions
Full size table

4.3 Optimization of Bayesian hyperparameters

In deep learning, hyperparameters are the variables that determine the network's structure and training method. Hyperparameters have a significant effect on the performance of the model, so they must be chosen carefully before training of the model. Key hyperparameters include: the number of the LSTM layer, the number of the LSTM cells, the number of neurons in each LSTM cell, the learning rate, the loss function, the batch size, and the dropout ratio. In this study, the number of LSTM cells is set to 10, because the sequence length of 1 s would produce 10 data points at the 10 Hz sampling rate. The cosine annealing learning rate method is used, which can reduce the learning rate according to a half cosine curve. This allows for relatively larger weight changes at the beginning of the learning process as well as small changes or fine-tuning towards the end of the learning process, thereby achieving better performance. The initial learning rate is set to 0.01. The cross-entropy loss function is selected, as it is commonly used for the classification problem. The remaining hyperparameters (i.e., the number of LSTM layer, the number of neurons, the batch size and the dropout ratio) are determined by a Bayesian optimization process (Brochu et al., 2010; Snoek et al., 2012).

Bayesian optimization works by constructing a posterior distribution of functions (Gaussian process) that best describes the target function (i.e., the function of validation accuracy with different hyperparameter combinations in this study). With increased number of hyperparameter combinations observed, the posterior distribution would be improved and the algorithm would become more certain. From this process, hyperparameter combination should be picked for the next observation. This process is designed to minimize the number of steps required for finding a combination of hyperparameters that is close to the optimal combination. The candidate set of each hyperparameter to be optimized is as follows:

  • The number of LSTM layers ∈ [1, 2, 3, 4, 5, 6]

  • The number of neurons in each LSTM cell ∈ [4, 8, 16, 32, 64, 128, 256]

  • Dropout ratio ∈ [0, 0.1, 0.2, 0.3, 0.4, 0.5]

  • Batch size ∈ [64, 128, 256, 512, 1024, 2048]

There are 1,512 hyperparameter combinations in total. Figure 6 shows the observed highest validation accuracy over iterations. The validation accuracy is improved in each round of iteration, and converge is observed at the 50th step.

Fig. 6
figure 6

Best observed results over iterations

Full size image

The optimum values of LSTM layers, neurons, dropout rate and batch size are 4, 256, 0.1 and 256, respectively. The optimization results are reasonable, as deeper (more layers) and wider (more neurons) models tend to have improved the performance of capturing the features hidden in sequences and yielded better performance. The dropout method can randomly ignore a certain percentage of neurons (dropout rate) in each layer during training, thus avoiding over-fitting and making the model more robust. The confusion matrix is shown in Table 3. Almost all pavements are correctly classified, with only 15 dry pavements misclassified as snowy or icy pavements and only 2 icy pavements misclassified as dry pavements. So, the classification accuracy of each pavement condition is close to 100%. The trained LSTM model can then be used for detection and classification of pavements based on the extent of slippery degrees.

Table 3 Confusion matrix of the best LSTM model on validation set
Full size table

4.4 Model prediction

The above results are all acquired based on the simulated CV data. However, in the field V2I applications, the readings of vehicle BSM data usually contain uncorrelated noises or abnormal values. Therefore, data cleaning should be implemented to reduce such noises, so as to improve the quality of data. The processed data can then be used for classification by the pre-trained model. The pseudo-code for pavement classification by slippery level is shown in Algorithm 1 .

5 The proposed warning model

The location of the detected slippery pavement is indicated in spatial coordinates in the CV data sequence. Figure 7 shows the map of pavements' slippery conditions along different road crossing sections.

Fig. 7
figure 7

Mapping of pavements' slippery conditions at each intersection during the 2710th ~ 2720th simulation seconds

Full size image

A warning model is developed to generate alerts to the approaching CVs based on the detection results. The pseudo-code for the warning model is summarized in Algorithm 2. .

The detected snowy and icy pavements by the trained LSTM classifier are likely to include misclassified results (false-positive samples); especially, some dry pavements may be misclassified as snowy or icy pavements, thus causing false warnings to be broadcasted. A clustering algorithm is used to mitigate the problem of false warnings. As shown in Line 2 of Algorithm 2, at each intersection, all detected snowy and icy pavements are clustered according to road segments. For example, for a four-way intersection (No. 1, 2, 3, and 5), there are four road segments representing four directions; and for a T-shaped intersection (No. 4 and 6), there are three road segments. The frequency of detected snowy and icy pavements on each road segment is counted. If the detected frequency of a road segment is lower than the predefined thresholds (dsnow and dicy), no warning will be broadcast. If both frequencies of detected snowy and icy pavements exceed the thresholds, warnings will be broadcast based on the dominant type of detected pavements. The warning will last for a certain period of time (Twarn), so as to ensure that the following CVs can receive the warning, thus avoiding potential accidents. In other words, the warning will stop if no snowy or icy pavements are detected for Twarn. The warning messages are broadcast at 1 Hz, in order to ensure the timeliness of delivering the information on road conditions. Based on sensitivity analyses, the optimal values of dsnow, dicy and Twarn are determined to be 5 (times), 5 (times) and 5 (minutes), respectively, in this study. In the field applications, these parameters can be adjusted according to traffic factors and weather conditions.

The warning model is tested and evaluated with VISSIM by generating new datasets under different environments. Each slippery condition scenario of pavements is simulated for 3 h and classified with the trained LSTM model. The confusion matrix of the warnings broadcast by the system is shown in Table 4: The first row indicates that all of the dry pavements are correctly recognized by LSTM, without any misclassifications and false warnings sent. Similarly, 99.06% of snowy pavements are correctly recognized, while only 0.94% of them are wrongly recognized as dry pavements, causing no warnings to have been broadcast promptly. The third row shows that 98.02% of icy pavements are correctly classified, while 1.98% of them are wrongly classified as snowy pavements, causing false warnings to have been broadcast; no icy pavement is classified as non-slippery dry pavement. Overall, the model has achieved high accuracy in detecting different slippery conditions of pavements.

Table 4 Confusion matrix of the warnings broadcast by the system
Full size table

6 Evaluation of safety improvement

The safety improvement acquired from the detection of slippery road conditions is evaluated by comparing the driving behaviors of CVs and normal vehicles. It is assumed that the drivers of CVs receiving the warnings will change the vehicle operation by taking the precautions as described in Table 1 (i.e., to reduce vehicle speeds and increase the following distances), while the drivers of the normal vehicles will keep their then-current operation, since they have not received any warning messages. The proportion is 50% for each of the vehicle types. The surrogate safety measures (SSMs) approach, one of the most widely used methods for identifying potential threats, such as rear-end collision, is used to identify the relationship between driving behaviors and collision risks (Yang, 2012). Three metrics, i.e., time to collision (TTC), modified time to collision (MTTC) and deceleration rate to avoid a crash (DRAC), are used to represent potential collision risks, which can be calculated with Eqs. 6–8 (Cooper & Ferguson, 1976; Ozbay et al., 2008; van der Horst & Hogema, 1993):

$${\text{TTC}} = \left\{ \begin{array}{ll} \frac{d}{{v_{f} - v_{l} }}, & v_{f} > v_{l}\\ \infty, & \, v_{f} \le v_{l} \\ \end{array} \right.$$
(6)
$${\text{MTTC}} = \frac{{ - \Delta v \pm \sqrt {\Delta v^{2} + 2\Delta ad} }}{\Delta a}$$
(7)
$${\text{DRAC}} = \left\{ \begin{array}{ll} \frac{{\left( {v_{f} - v_{l} } \right)^{2} }}{d}, & \, v_{f} > v_{l} \\ 0, & \,\, v_{f} \le v_{l} \\ \end{array} \right.$$
(8)

where d indicates the relative space gap between two vehicles; vf indicates the then-current speed of the following vehicle; vl indicates the then-current speed of the leading vehicle; Δv indicates the relative speed between two vehicles; and Δa indicates the relative acceleration between two vehicles.

For each simulation vehicle, all the three metrics above can be calculated to represent the potential driving risk at each simulation second. The recommended thresholds of TTC, MTTC and DRAC for identifying potential collisions are 2.4 s, 3.5 s and 4 m/s2, respectively (Archer, 2005; Hogema & Janssen, 1996). In other words, for a simulation vehicle at a specific simulation time, if its TTC is less than 2.4 s or its MTTC is less than 3.5 s or its DRAC is higher than 4 m/s2, the driving behavior then will be risky and may lead to potential collision. By counting the number of risky driving behaviors in each road segment, the safety characteristic of intersections can be evaluated. The histograms of TTC, MTTC and DRAC of CVs and normal vehicles under different pavement conditions at all simulation intersections are shown in Fig. 8, where the red dotted vertical lines indicate the thresholds. It should be noted that under the dry pavements, the CVs and normal vehicles demonstrate the same SSMs distribution, since no warnings are broadcasted; on the other hand, under snowy and icy pavements, the potential collisions observed by CVs are much less than by normal vehicles. TTC, MTTC and DRAC are decreased by 37%, 30% and 49% under snowy pavement conditions, and by 77%, 66% and 99% under icy pavement conditions. This indicates that the warnings of slippery road conditions delivered to the approaching vehicles have significantly reduced potential collisions and improved the safety.

Fig. 8
figure 8

Histograms of SSMs of CVs and normal vehicles under different pavement conditions

Full size image

7 Conclusion

An innovative CV-based system is proposed for slippery pavement detection, purposed to provide real-time warning of pavements' slippery conditions. VISSIM with a CV add-on module is used to generate a traffic network and CV data based on a road section of the SPMD program in Ann Arbor. The vehicle kinematics is acquired based on the vehicle tire-pavement interaction model, and the pavements are classified into three major categories, i.e., dry, snowy and icy pavements, based on their extent of slipperiness. A deep learning network, LSTM, is developed to classify pavements' slippery conditions based on the vehicle kinetics recorded in the BSM of CV data. The effect of different hyperparameters on the classification performance is investigated, and the optimized set of hyperparameters for the LSTM model is determined. The testing results show that the LSTM classifier achieves very high classification accuracy in determining the slippery conditions of pavements, with the detection accuracy reaching 100%, 99.06% and 98.02% for dry, snowy and icy pavements, respectively. The benefits of slippery road detection are analyzed by assuming that drivers of CVs would change their driving behaviors when receiving warning messages on slippery roads. Compared with normal vehicles with no pavement slippery information received, CVs receiving road slippery messages have demonstrated significant improvements in safety features. Integrated with advanced deep learning models, the CV-based technology is promising to map pavements' slippery conditions in real time and improve road users' safety by delivering warning information to drivers. This function is particularly important under adverse weather conditions.