Naturalistic Driving Study Research Articles

Detection of Driver Distraction in the Australian Naturalistic Driving Study Videos Using Pre-Trained Models and Transfer Learning

Detection of Driver Distraction in the Australian Naturalistic Driving Study Videos Using Pre-Trained Models and Transfer Learning

Transfollower: Long-Sequence Car-Following Trajectory Prediction Through Transformer

Car-following refers to a control process in which the following vehicle (FV) tries to keep a safe distance between itself and the lead vehicle (LV) by adjusting its acceleration in response to the actions of the vehicle ahead. The corresponding car-following models, which describe how one vehicle follows another vehicle in the traffic flow, form the cornerstone for microscopic traffic simulation and intelligent vehicle development. One major motivation of car-following models is to replicate human drivers' longitudinal driving trajectories. To model the long-term dependency of future actions on historical driving situations, we developed a long-sequence car-following trajectory prediction model based on the attention-based Transformer model. The model follows a general format of encoder-decoder architecture. The encoder takes historical speed and spacing data as inputs and forms a mixed representation of historical driving context using multi-head self-attention. The decoder takes the future LV speed profile as input and outputs the predicted future FV speed profile in a generative way (instead of an auto-regressive way, avoiding compounding errors). Through cross-attention between encoder and decoder, the decoder learns to build a connection between historical driving and future LV speed, based on which a prediction of future FV speed can be obtained. We train and test our model with 112,597 real-world car-following events extracted from the Shanghai Naturalistic Driving Study (SH-NDS). Results show that the model outperforms the traditional intelligent driver model (IDM), a fully connected neural network model, and a long short-term memory (LSTM) based model in terms of long-sequence trajectory prediction accuracy. We also visualized the self-attention and cross-attention heatmaps to explain how the model derives its predictions.

Stacking‐based ensemble learning method for the recognition of the preceding vehicle lane‐changing manoeuvre: A naturalistic driving study on the highway

The recognition of the preceding vehicle lane-changing manoeuvre (LCM) is essential for improving the rationality and safety of the decision-making of driverless vehicles. However, traditional recognition researches for preceding vehicle LCM are generally characterized by problems such as low recognition accuracy and poor research foundation. In response to these problems, this paper carried out naturalistic driving study (NDS) on the highway, from which a large amount of on-road data consisting of lane-keeping (LK) and lane-changing left and right (LCL, LCR) manoeuvres were collected. A stacking-based ensemble learning method for the recognition of the LCM of the preceding vehicle, which integrates the random forest, support vector machine, long and short-term memory network (LSTM), and bi-directional LSTM based on attention mechanism (AT-Bi-LSTM) algorithms, is proposed. Compared with traditional machine learning methods, the proposed method exhibits greater advantages in terms of its recognition accuracy. It is particularly of note that the recognition accuracy of the model at 0.4 s and 0.8 s after the LCM reached 90.77% and 95.54%, respectively. The study reported is of great significance for the construction of more intelligent vehicle-vehicle collaboration and the promotion of the industrial applications of intelligent vehicle technology.

Driver Behavior on Freeway Entrance Ramp Terminals

AbstractThis paper studied driver behavior on freeway entrance ramp terminals using trip data from the SHRP-2 Naturalistic Driving Study (NDS) database from two different US states. The study inclu...

Modeling Car-Following Behavior on Freeways Considering Driving Style

To build more accurate and realistic freeway car-following models, driving characteristics specific to freeway car following should be considered. This study, therefore, analyzed three car-following models calibrated for different driving styles. A total of 5,900 freeway car-following events were extracted from 161,055 km of driving data collected in the Shanghai Naturalistic Driving Study (SH-NDS) database. Based on the fuzzy inference system built in this study, these car-following events were categorized as representing one of two styles: nonaggressive or aggressive. The two driving styles were visualized by using the t-distributed stochastic neighbor embedding (t-SNE) algorithm. The Gipps, Wiedemann, and intelligent driver model (IDM) car-following models were calibrated and validated for each driving style group. Using a genetic algorithm to analyze the calibrated parameters of the investigated car-following models, it was found that the model parameter values were related to driving style. When their performances were evaluated, results showed that the IDM performed best. The nonaggressive IDM and the aggressive IDM were used to simulate the car-following scenarios based on the same leading vehicle trajectories. The t-test and the F-test results showed that regarding both time gap and spacing gap, the differences of mean and variance are significant between aggressive and nonaggressive styles in nearly all of the simulated car-following scenarios. The mean spacing gap (nonaggressive: 38 m; aggressive: 30 m) and time gap (nonaggressive: 1.7 s; aggressive: 1.4 s) obtained from modeled car-following scenarios could be used directly in simulation software to show the characteristics of different driving styles.

Aggressive driving behavior prediction considering driver’s intention based on multivariate-temporal feature data

Aggressive driving behavior is mainly motivated by the intention of the driver; therefore, the underlying intention of behavior should be considered in investigating aggressive driving behavior. However, existing aggressive driving behavior prediction methods are not advanced in a compelling of characterizing the driver's intention among a large set of attributes and describing the random process among time-varying transversion. To address this, this paper proposes a prediction method, which is structured with a Hidden Markov Model (HMM) and attention-based Long Short-Term Memory (LSTM) Network. HMM is applied to extract the driver's intention which leads to aggressive driving behavior; attention-based LSTM networks are applied in the multivariate-temporal aggressive driving behavior prediction. The method input uses panel data which contains observations about different cross-sections across time. In the case study, the model was trained based on the Shanghai Naturalistic Driving Study data. After comparing with other deep learning methods and normal LSTM, results show the proposed method provides good performance for aggressive driving behavior prediction (Mean of Accuracy = 80%), especially with the 2-sec time interval applied (Training Accuracy = 82% and Validation Accuracy = 84%). Also, the result shows that the attention mechanism can improve the result's interpretability, and using the driver’s intentions as input can enhance the model accuracy. This method for predicting aggressive driving behavior that combines driver's intention, variable contribution sorting, and time-series processing. This method can be used in real-world applications for improving driving safety with the applications in the Advanced Driver Assistance Systems.

In-depth analysis of crash contributing factors and potential ADAS interventions among at-risk drivers using the SHRP 2 naturalistic driving study

Objective Motor vehicle crashes remain a significant problem. Advanced driver assistance systems (ADAS) have the potential to reduce crash incidence and severity, but their optimization requires a comprehensive understanding of driver-specific errors and environmental hazards in real-world crash scenarios. Therefore, the objectives of this study were to quantify contributing factors using the Strategic Highway Research Program 2 (SHRP 2) Naturalistic Driving Study (NDS), identify potential ADAS interventions, and make suggestions to optimize ADAS for real-world crash scenarios. Methods A subset of the SHRP 2 NDS consisting of at-fault crashes (n = 369) among teens (16–19 yrs), young adults (20–24 yrs), adults (35–54 yrs) and older adults (70+ yrs) were reviewed to identify contributing factors and potential ADAS interventions. Contributing factors were classified according to National Motor Vehicle Crash Causation Survey pre-crash assessment variable elements. A single critical factor was selected among the contributing factors for each crash. Case reviews with a multidisciplinary panel of industry experts were conducted to develop suggestions for ADAS optimization. Critical factors were compared across at-risk driving groups, gender, and incident type using chi-square statistics and multinomial logistic regression. Results Driver error was the critical factor in 94% of crashes. Recognition error (56%), including internal distraction and inadequate surveillance, was the most common driver error sub-type. Teens and young adults exhibited greater decision errors compared to older adults (p < 0.01). Older adults exhibited greater performance errors (p < 0.05) compared to teens and young adults. Automatic emergency braking (AEB) had the greatest potential to mitigate crashes (48%), followed by vehicle-to-vehicle communication (38%) and driver monitoring (24%). ADAS suggestions for optimization included (1) implementing adaptive forward collision warning, AEB, high-speed warning, and curve-speed warning to account for road surface conditions (2) ensuring detection of nonstandard road objects, (3) vehicle-to-vehicle communication alerting drivers to cross-traffic, (4) vehicle-to-infrastructure communication alerting drivers to the presence of pedestrians in crosswalks, and (5) optimizing lane keeping assist for end-departures and pedal confusion. Conclusions These data provide stakeholders with a comprehensive understanding of critical factors among at-risk drivers as well as suggestions for ADAS improvements based on naturalistic data. Such data can be used to optimize ADAS for driver-specific errors and help develop more robust vehicle test procedures.

Analysis of Headway and Speed Based on Driver Characteristics and Work Zone Configurations Using Naturalistic Driving Study Data

The objective of this paper is to analyze headway and speed distribution based on driver characteristics and work zone (WZ) configurations by utilizing Naturalistic Driving Study (NDS) data. The NDS database provides a unique opportunity to study car-following behaviors for different driver types in various WZ configurations, which cannot be achieved from traditional field data collection. The complete NDS WZ trip data of 200 traversals and 103 individuals, including time-series data, forward-view videos, radar data, and driver characteristics, was collected at four WZ configurations, which encompasses nearly 1,100 vehicle miles traveled, 19 vehicle hours driven, and over 675,000 data points at 0.1 s intervals. First, the time headway selections were analyzed with driver characteristics such as the driver’s gender, age group, and risk perceptions to develop the headway selection table. Further, the speed profiles for different WZ configurations were established to explore the speed distribution and speed change. The best-fitted curves of time headway and speed distributions were estimated by the generalized additive model (GAM). The change point detection method was used to identify where significant changes in mean and variance of speeds occur. The results concluded that NDS data can be used to improve car-following models at WZs that have been implemented in current WZ planning and simulation tools by considering different headway distributions based on driver characteristics and their speed profiles while traversing the entire WZ.

Differences in Near-Crash Risk by Types of Distraction: A Comparison of Trends between Freeways and Two-Lane Highways using Naturalistic Driving Data

Distracted driving is among the leading causes of motor vehicle crashes in the United States, though the magnitude of this problem is difficult to quantify given limitations of police-reported crash data. This study leveraged data from the second Strategic Highway Research Program Naturalistic Driving Study to gain important insights into the risks posed by driver distraction on both freeways and two-lane highways. More than 50 types of secondary tasks were aggregated into ten distraction type categories and mixed-effects logistic regression models were estimated to discern how the risks of near-crash events varied by distraction type while controlling for the effects of driver, roadway, and traffic characteristics. In general, the types of distractions that created the most pronounced risks were those that introduced a combination of cognitive, visual, and manual distractions. For example, drivers who used cell phones were subject to higher risks and these risks tended to be most pronounced when both visual and manual distractions were involved. Likewise, risks tended to be highest when drivers reached for other objects inside the vehicle, engaged in personal hygiene-related activities, or focused on activities occurring outside of the driving environment. Although the same factors tended to increase near-crash risk on both types of facilities, the impacts of several factors tended to be more pronounced on two-lane highways where interaction with other vehicles occurred more frequently. From a policy standpoint, the results of this study provide further motivation for more aggressive legislation and enforcement of distracted driving.

Quantifying the effect of roadway, driver, vehicle, and location characteristics on the frequency of longitudinal and lateral accelerations

The purpose of this study is to understand and quantify the simultaneous effects of roadway speed category, driver age, driver gender, vehicle class, and location on the rates of longitudinal and lateral acceleration epochs. The rate of usual as well as harsh acceleration epochs are used to extract insights on driving risk and driver comfort preferences. However, an analysis of acceleration rates at multiple thresholds incorporating various effects while using a large-scale and diverse dataset is missing. This analysis will fill this research gap. Data from the 2nd Strategic Highway Research Program Naturalistic Driving Study (SHRP2 NDS) was used for this analysis. The rate of occurrence of acceleration epochs was modeled using negative binomial distribution based generalized linear mixed effect models. Roadway speed category, driver age, driver gender, vehicle class, and location were used as the fixed effects and the driver identifier was used as the random effect. Incidence rate ratios were then calculated to compare subcategories of each fixed effect. Roadway speed category has the strongest effect on longitudinal and lateral accelerations of all magnitudes. Acceleration epoch rates consistently decrease as the roadway speed category increases. The difference in the rates depends on the threshold and is up to three orders of magnitude. Driver age is another significant factor with clear trends for longitudinal and lateral acceleration epochs. Younger and older drivers experience higher rates of longitudinal accelerations and decelerations. However, the rate of lateral accelerations consistently decreases with age. Vehicle class also has a significant effect on the rate of harsh accelerations with minivans consistently experiencing lower rates.

Development and Comparative Analysis of Advanced Deep Learning Techniques for Crash Prediction in Advanced Driver Support Systems

Motor vehicle crashes claimed 38,800 lives and caused 4.4 million injuries in 2019 alone. Studies have shown that 94% of these crashes are because of driver errors. Such a huge contribution of driver errors to crashes indicates that efforts at improving safety should be directed toward both vehicles and drivers through advanced driver assistance systems (ADAS) and vehicular technologies. This study investigates the potential that real-time driver behavior data collected through vehicular technologies offer to predict crashes, as the first line of defense to avoid them. Three deep learning models were developed including multilayer perceptron neural networks (MLP-NN), long-short-term memory networks (LSTMN), and convolutional neural networks (CNN) using vehicle kinematics time series data extracted from the Second Strategic Highway Research Program Naturalistic Driving Study (SHRP2 NDS) dataset. The study builds on the hypothesis that crashes are preceded by turbulences that take place over time (turbulence horizon). If these turbulences are detected promptly they can help predict and avoid crashes. Several values were tested for the turbulence horizon and the prediction horizon (how long before the crash impact it can be predicted) to identify the optimal values. The results showed that the CNN model can predict all crashes with a 100% accuracy and zero false alarms 3 s before the crash impact time when a 6-s turbulence horizon is used. This outstanding performance demonstrates the developed model is a promising tool for implementation in ADAS.

Factors modifying the likelihood of speeding behaviors based on naturalistic driving data

Speeding behaviors are quite common and are known to affect the risk and severity outcomes of vehicular crashes. Naturalistic driving data allows for the direct observation of speeding and the development of evidence-based causal structures for this behavior. This limits biases associated with self-reported speeding prevalence, allowing for more precise speeding measures than post-crash investigations and for the evaluation of driver attributes associated with speeding across a wide variety of locations and road types. Data from the Second Strategic Highway Research Program Naturalistic Driving Study were used to identify speeding events that were aggregated based on their duration and degree of speeding above the posted speed limit. The events were summarized as a likelihood of speeding metric for each trip in the dataset. These likelihood of speeding measurements were also aggregated across drivers and compared, using a beta binomial regression model, to driver questionnaire answers that addressed drivers’ perception of their speeding as well as different driver-specific factors that are suspected of having an influence on speeding behaviors. Results showed that, consistent with past studies, age and gender significantly influenced the likelihood of speeding. For age, the odds of speeding exhibited a significant downward trend across increasing age groups; 16–24 year olds exhibited odds of speeding that were 1.5 times the odds of drivers that were 80 or more years old. For gender, males exhibited larger odds of speeding than females (1.1 times larger). In addition, the odds of speeding were larger at lower posted speed limits. The odds of speeding in 10–20 mph zones were 9.5 times the odds of zones with speed limits greater than 60 mph, implying that drivers may be unaware of the risks associated with speeding in low speed limit areas. Several questionnaire answers related to drivers’ perception of speeding were also predictive of the likelihood of speeding.

Driver influence on vehicle trajectory prediction

Drivers continually interact with other road users and use information from the road environment to make decisions to control their vehicle. A clear understanding of different parameters impacting this interaction can provide us with a new design approach for a more effective driver assistance system - a personalised trajectory prediction system. This paper highlights the influential factors on trajectory prediction system performance by (i) identifying driver behaviours impacting the trajectory prediction system; and (ii) analysing other contributing factors such as traffic density, secondary task, gender and age group. To explore the most influential contributing factors, we first train an interaction-aware trajectory prediction system using time-series data derived from the Second Strategic Highway Research Program (SHRP2) Naturalistic Driving Study (NDS). Prediction error is then analysed based on driver characteristics such as driver profile which is subjectively measured through self-reported questions, and driving performance which is based on evaluation of time-series information such as speed, acceleration, jerk, time, and space headway. The results show that prediction error significantly increased in the scenarios where the driver engaged in risky behaviour. Analysis shows that trajectory prediction system performance is also affected by factors such as traffic density, engagement in secondary tasks, driver gender and age group. We show that the driver profile, which is subjectively measured using self-reported questionnaires, is not as significant as the driving performance information, which is objectively measured and extracted during each specific driving scenario.

Built environment, driving errors and violations, and crashes in naturalistic driving environment

Driving errors and violations are highly relevant to the safe systems approach as human errors tend to be a predominant cause of crash occurrence. In this study, we harness highly detailed pre-crash Naturalistic Driving Study (NDS) data 1) to understand errors and violations in crash, near-crash, and baseline (no event) driving situations, and 2) to explore pathways that lead to crashes in diverse built environments by applying rigorous modeling techniques. The “locality” factor in the NDS data provides information on various types of roadway and environmental surroundings that could influence traffic flow when a precipitating event is observed. Coded by the data reductionists, this variable is used to quantify the associations of diverse environments with crash outcomes both directly and indirectly through mediating driving errors and violations. While the most prevalent errors in crashes were recognition errors such as failing to recognize a situation (39 %) and decision errors such as not braking to avoid a hazard (34 %), performance errors such as poor lateral or longitudinal control or weak judgement (8 %) were most strongly correlated with crash occurrence. Path analysis uncovered direct and indirect relationships between key built-environment factors, errors and violations, and crash propensity. Possibly due to their complexity for drivers, urban environments are associated with higher chances of crashes (by 6.44 %). They can also induce more recognition errors which correlate with an even higher chances of crashes (by 2.16 % with the “total effect” amounting to 8.60 %). Similar statistically significant mediating contributions of recognition errors and decision errors near school zones, business or industrial areas, and moderate residential areas were also observed. From practical applications standpoint, multiple vehicle technologies (e.g., collision warning systems, cruise control, and lane tracking system) and built-environment (roadway) changes have the potential to reduce driving errors and violations which are discussed in the paper.

Quantification of Rear-End Crash Risk and Analysis of Its Influencing Factors Based on a New Surrogate Safety Measure

Traditional surrogate measures of safety (SMoS) cannot fully consider the crash mechanism or fail to reflect the crash probability and crash severity at the same time. In addition, driving risks are constantly changing with driver’s personal driving characteristics and environmental factors. Considering the heterogeneity of drivers, to study the impact of behavioral characteristics and environmental characteristics on the rear-end crash risk is essential to ensure driving safety. In this study, 16,905 car-following events were identified and extracted from Shanghai Naturalistic Driving Study (SH-NDS). A new SMoS, named rear-end crash risk index (RCRI), was then proposed to quantify rear-end crash risk. Based on this measure, a risk comparative analysis was conducted to investigate the impact of factors from different facets in terms of weather, temporal variables, and traffic conditions. Then, a mixed-effects linear regression model was applied to clarify the relationship between rear-end crash risk and its influencing factors. Results show that RCRI can reflect the dynamic changes of rear-end crash risk and can be applied to any car-following scenarios. The comparative analysis indicates that high traffic density, workdays, and morning peaks lead to higher risks. Moreover, results from the mixed-effects linear regression model suggest that driving characteristics, traffic density, day-of-week (workday vs. holiday), and time-of-day (peak hours vs. off-peak hours) had significant effects on driving risks. This study provides a new surrogate safety measure that can better identify rear-end crash risks in a more reliable way and can be applied to real-time crash risk prediction in driver assistance systems. In addition, the results of this study can be used to provide a theoretical basis for the formulation of traffic management strategies to improve driving safety.

Understanding Gap Acceptance Behavior at Unsignalized Intersections using Naturalistic Driving Study Data

Even though drivers disregarding a stop sign is widely considered a major contributing factor for crashes at unsignalized intersections, an equally important problem that leads to severe crashes at such locations is misjudgment of gaps. This paper presents the results of an effort to fully understand gap acceptance behavior at unsignalized intersections using SHPR2 Naturalistic Driving Study data. The paper focuses on the findings of two research activities: the identification of critical gaps for common traffic/roadway scenarios at unsignalized intersections, and the investigation of significant factors affecting driver gap acceptance behaviors at such intersections. The study used multiple statistical and machine learning methods, allowing a comprehensive understanding of gap acceptance behavior while demonstrating the advantages of each method. Overall, the study showed an average critical gap of 5.25 s for right-turn and 6.19 s for left-turn movements. Although a variety of factors affected gap acceptance behaviors, gap size, wait time, major-road traffic volume, and how frequently the driver drives annually were examples of the most significant.

An integrated methodology for real-time driving risk status prediction using naturalistic driving data

Real-time driving risk status prediction is critical for developing proactive traffic intervention strategies and enhance driving safety. However, the optimal observation time window length and prediction time window length, which should be the prerequisite for the timeliness and accuracy of real-time driving risk status prediction model, have been rarely explored in previous studies. In this study, a methodology which integrates driving risk status identification, rolling time window-based feature extraction, real-time driving risk status prediction and driving risk influencing factors analysis was proposed to accurately evaluate and predict real-time driving risk status. The methodology was tested based on 1,440 car-following events from Shanghai Naturalistic Driving Study. Results show that four driving risk statuses (safe, low-risk, median-risk and high-risk) are most appropriate to establish risk labelling criteria. In addition, results from driving risk status prediction show that when the observation time window length is 0.5 s, the accuracy rate of predicting medium-risk or high-risk status occurring in the next 0.7 s is higher than 85 % using multi-layer perceptron model. Meanwhile, the results from the analysis of influencing factors show that the input variables related to the risk status score higher in the ranking of feature importance. A part from that, speed difference, headway distance, speed and acceleration are still important in predicting driving risk status. The proposed methods in this paper can be applied in connected and autonomous vehicle (CAV) to reduce driver cognitive workload and hence improve driving safety fed with naturalistic driving data collected using in-vehicle systems.

Using SHRP2 NDS data to examine infrastructure and other factors contributing to older driver crashes during left turns at signalized intersections

Drivers age 65 and over have higher rates of crashes and crash-related fatalities than other adult drivers and are especially over-represented in crashes during left turns at intersections. This research investigated the use of SHRP2 Naturalistic Driving Study (NDS) data to assess infrastructure and other factors contributing to left turn crashes at signalized intersections, and how to improve older driver safety during such turns. NDS data for trips involving signalized intersections and crash or near-crash events were obtained for two driver age groups: drivers age 65 and over (older drivers) and a sample of drivers age 30−49, along with NDS pre-screening and questionnaire data. Video scoring of all trips was performed to collect additional information on intersection and trip conditions. To identify the most influential factors of crash risk during left turns at signalized intersections, machine learning and regression models were used. The results found that in the obtained NDS dataset, there was a relatively small volume of crashes during left turns at signalized intersections. Further, model results found the statistically significant variables of crash risk for older drivers were associated more with health and cognitive factors rather than the infrastructure or design of the intersections. The results suggest that a study using only SHRP2 NDS data will not lead to definitive findings or recommendations for infrastructure changes to increase safety for older drivers at signalized intersections and during left turns. Moreover, the findings of this study indicates the need to consider other data sources and data collection methods to address this critical literature gap in older driver safety.

Crash Factor Analysis in Intersection-Related Crashes Using SHRP 2 Naturalistic Driving Study Data

&lt;div class="section abstract"&gt;&lt;div class="htmlview paragraph"&gt;Intersections have a high risk of vehicle-to-vehicle conflicts because of the overlapping traffic flow from multiple roads. To understand the factors contributing to the crashes, this study examines the common characteristics in intersection-related crash and near- crash events, such as the existence of traffic control devices, the driver at fault, and occurrence of visual obstructions. The descriptive data of the crash and near-crash events recorded in the Second Strategic Highway Research Program Naturalistic Driving Study (SHRP 2 NDS) database is used in categorization and statistical analysis in this study. First, the events are divided into seven categories based on trajectories of the conflicting vehicles. The categorization provides the basis for in-depth analysis of crash-contributing factors in specific confliction patterns. Subsequently, descriptive statistics are used to portray each of the categories. The severity of the categories is determined based of the frequency of occurrences of the crashes and near-crashes. Factors contributing to each crash category are then examined based on their frequency of occurrence with the categories. This study reveals the severity of different intersection-related crash patterns, as well as the effects of some crash-contributing factors. The information presented in this paper has the potential to guide investigators interested in developing simulation and test track scenarios to evaluate vehicles equipped with automated driving systems (ADS).&lt;/div&gt;&lt;/div&gt;

Development of a Novel Convolutional Neural Network Architecture Named RoadweatherNet for Trajectory-Level Weather Detection using SHRP2 Naturalistic Driving Data

Driver performances could be significantly impaired in adverse weather because of poor visibility and slippery roadways. Therefore, providing drivers with accurate weather information in real time is vital for safe driving. The state-of-practice of collecting roadway weather information is based on weather stations, which are expensive and cannot provide trajectory-level weather information. Therefore, the primary objective of this study was to develop an affordable detection system capable of providing trajectory-level weather information at the road surface level in real-time. This study utilized the Strategic Highway Research Program 2 Naturalistic Driving Study video data combined with a promising machine learning technique, called convolutional neural network (CNN), to develop a weather detection model with seven weather categories: clear, light rain, heavy rain, light snow, heavy snow, distant fog, and near fog. A novel CNN architecture, named RoadweatherNet, was carefully crafted to achieve the weather detection task. The evaluation results based on a test dataset revealed that RoadweatherNet can provide excellent performance in detecting weather conditions with an overall accuracy of 93%. The performance of RoadweatherNet was also compared with six pre-trained CNN models, namely, AlexNet, ResNet18, ResNet50, GoogLeNet, ShuffleNet, and SqueezeNet, which showed that RoadweatherNet can provide nearly identical performance with a significant reduction in training time. The proposed weather detection model is cost-efficient and requires less computational power; therefore, it can be made widely available mainly owing to the recent thriving of smartphone cameras and can be used to expand and update the current weather-based variable speed limit systems.