How AI and Connected Vehicles Are Making Winter Roads Safer

The study categorizes pavement surfaces into three distinct types based on their friction coefficients: dry, snowy, and icy. Among these, dry surfaces exhibit the least slipperiness, whereas icy surfaces demonstrate the highest degree of slipperiness. The process of slippery detection can be effectively framed as a classification problem, which is well-suited for deep learning methodologies. Statistical data derived from numerous road surveys provide detailed ranges for friction coefficients associated with various surface conditions: dry surfaces exhibit coefficients ranging from 0.85 to 1, snowy surfaces range from 0.2 to 0.3, and icy surfaces range from 0.05 to 0.2. These coefficients have significant implications for driving safety and inform the corresponding preventive measures applicable to different pavement conditions.

The preparation process for the dataset intended for slippery detection comprises several crucial steps. Initially, the data is standardized to achieve a mean of 0 and a standard deviation of 1. After that, the dataset is divided into 1-second frames, which aligns with the 10 Hz sampling rate, resulting in 10 data points for each frame. Subsequently, each sequence is tagged based on pavement conditions, classifying them as either dry, snowy, or icy. To address potential class imbalance, methods like under-sampling and over-sampling are implemented, ensuring each category contains 282,197 sequences. In the end, the balanced dataset is divided into an 80% training set and a 20% validation set, employing stratified sampling to preserve the class distribution. The training set is used to train the LSTM model, while the validation set is allocated for hyperparameter optimization.

In deep learning, hyperparameters play a crucial role in determining model performance and must be selected with care before training begins. Important hyperparameters include the number of LSTM layers, LSTM cells, neurons per cell, learning rate, loss function, batch size, and dropout ratio. For this study, the number of LSTM cells is fixed at 10, corresponding to a sampling rate of 10 Hz. The cosine annealing learning rate technique is applied, starting at an initial learning rate of 0.01, permitting larger initial adjustments followed by gradual fine-tuning. The cross-entropy loss function is employed for classification tasks. The other hyperparameters are refined using Bayesian optimization. The optimal settings for the LSTM layers, neurons, dropout rate, and batch size are 4, 256, 0.1, and 256, respectively. The optimization outcomes are justifiable, as deeper (more layers) and wider (more neurons) architectures generally enhance the ability to capture intricate features in sequences, resulting in superior performance.

The simulations were performed utilising the VISSIM traffic modelling software, which demonstrated that the LSTM model exhibits remarkable accuracy in detecting a range of pavement conditions. Specifically, the model achieved 100% accuracy for dry pavement, 99.06% for snowy conditions, and 98.02% for icy pavement. Upon identifying a hazardous road condition, the AI model activates a warning system that disseminates alerts to nearby connected vehicles (CVs).

Additionally, the study investigated the influence of these warnings on driver behaviour. It was observed that vehicles that adjusted their speed and maintained a greater following distance after receiving an alert experienced a significantly reduced number of near-collision events compared to those that did not receive any alerts. Key findings from the simulation data indicated that safety indicators such as time-to-collision (TTC), modified time-to-collision (MTTC), and deceleration rate to avoid a crash (DRAC) improved by 30-77% for connected vehicles in contrast to their non-connected counterparts. Furthermore, potential accident rates decreased by over 90% when drivers adhered to the recommended safety measures following an alert.

The integration of AI and connected vehicle technology is revolutionizing winter road safety. Utilizing real-time data and sophisticated machine learning methods, we can develop a proactive system that detects and helps prevent accidents on icy and snowy roads. As these technologies advance, we approach a future with enhanced road safety, making winter driving safer for all.