Statistics affirm that almost half of deaths in traffic accidents were vulnerable road users, such as pedestrians, cyclists, and motorcyclists. Despite the efforts in technological infrastructure and traffic policies, the number of victims remains high and beyond expectation. Recent research establishes that determining the causes of traffic accidents is not an easy task because their occurrence depends on one or many factors. Traffic accidents can be caused by, for instance, mechanical problems, adverse weather conditions, mental and physical fatigue, negligence, potholes in the road, among others. At present, the use of learning-based prediction models as mechanisms to reduce the number of traffic accidents is a reality. In that way, the success of prediction models depends mainly on how data from different sources can be integrated and correlated. This study aims to report models, algorithms, data sources, attributes, data collection services, driving simulators, evaluation metrics, percentages of data for training/validation/testing, and others. We found that the performance of a prediction model depends mainly on the quality of its data and a proper data split configuration. The use of real data predominates over data generated by simulators. This work made it possible to determine that future research must point to developing traffic accident prediction models that use deep learning. It must also focus on exploring and using data sources, such as driver data and light conditions, and solve issues related to this type of solution, such as high dimensionality in data and information imbalance.

A complex AI system must be worked offline because the training and execution phases are processed separately. This process often requires different computer resources due to the high model requirements. A limitation of this approach is the convoluted training process that needs to be repeated to obtain models with new data continuously incorporated into the knowledge base. Although the environment may be not static, it is crucial to dynamically train models by integrating new information during execution. In this article, artificial neural networks (ANNs) are developed to predict risk levels in traffic accidents with relatively simpler configurations than a deep learning (DL) model, which is more computationally intensive. The objective is to demonstrate that efficient, fast, and comparable results can be obtained using simple architectures such as that offered by the Radial Basis Function neural network (RBFNN). This work led to the generation of a driving dataset, which was subsequently validated for testing ANN models. The driving dataset simulated the dynamic approach by adding new data to the training on-the-fly, given the constant changes in the drivers’ data, vehicle information, environmental conditions, and traffic accidents. This study compares the processing time and performance of a Convolutional Neural Network (CNN), Random Forest (RF), Radial Basis Function (RBF), and Multilayer Perceptron (MLP), using evaluation metrics of accuracy, Specificity, and Sensitivity-recall to recommend an appropriate, simple, and fast ANN architecture that can be implemented in a secure alert traffic system that uses encrypted data.

Recently, the number of vehicles equipped with wireless connections has increased considerably. The impact of that growth in areas such as telecommunications, infotainment, and automatic driving is enormous. More and more drivers want to be part of a vehicular network, despite the implications or risks that, for instance, the openness of wireless communications, its dynamic topology, and its considerable size may bring. Undoubtedly, this trend is because of the benefits the vehicular network can offer. Generally, a vehicular network has two modes of communication (V2I and V2V). The advantage of V2I over V2V is roadside units’ high computational and transmission power, which assures the functioning of early warning and driving guidance services. This paper aims to discover the principal vulnerabilities and challenges in V2I communications, the tools and methods to mitigate those vulnerabilities, the evaluation metrics to measure the effectiveness of those tools and methods, and based on those metrics, the methods or tools that provide the best results. Researchers have identified the non-resistance to attacks, the regular updating and exposure of keys, and the high dependence on certification authorities as main vulnerabilities. Thus, the authors found schemes resistant to attacks, authentication schemes, privacy protection models, and intrusion detection and prevention systems. Of the solutions for providing security analyzed in this review, the authors determined that most of them use metrics such as computational cost and communication overhead to measure their performance. Additionally, they determined that the solutions that use emerging technologies such as fog/edge/cloud computing present better results than the rest. Finally, they established that the principal challenge in V2I communication is to protect and dispose of a safe and reliable communication channel to avoid adversaries taking control of the medium.

The problems with current driving datasets are their exclusivity to autonomous driving applications and their limited diversity in terms of sources of information and number of attributes. Thus, this paper presents a novel driving dataset that contains information from several heterogeneous sources and targets road traffic safety applications. We used an acquisition module based on software and hardware to collect information from a vehicle scanner and a health monitor. This module also consumes information from a weather web service and databases on traffic accidents and road geometric characteristics. For the acquisition sessions, drivers of different ages and genders drove vehicles on two routes at different day hours in different weather conditions. POLIDriving contains around 18 h of driving data, more than 61k observations, and 32 attributes. Unlike the other related datasets that include information on vehicle and road conditions, POLIDriving also includes information on the driver, weather conditions, traffic accidents, and road geometric characteristics. The dataset was tested in learning models to predict the risk levels of suffering a traffic accident. Hence, we built two learning models: Gradient Boosting Machine (GBM) and Multilayer Perceptron (MLP). GBM reached an accuracy value of 95.6%, and MLP reached an accuracy of 98.6%. Undoubtedly, POLIDriving will contribute greatly to the research on traffic accident prevention by providing a novel, numerous, and diverse driving dataset.

Due to the expansion of Artificial Intelligence (AI), especially Machine Learning (ML), it is more common to face confidentiality regulations about using sensitive data in learning models generally hosted in cloud environments. Confidentiality regulations such as HIPAA and GDPR seek to guarantee the confidentiality and privacy of personal information. Input and output data of a learning model may include sensitive data that must be protected. Adversaries could intercept and exploit this data to infer more sensitive data or even to determine the structure of the prediction model. To guarantee data privacy, one option could be encrypting data and making inferences over encrypted data. This strategy would be challenging for learning models that now must receive encrypted data, make inferences over encrypted data, and deliver encrypted data. To address this issue, this paper presents a privacy-preserving machine learning approach using Fully Homomorphic Encryption (FHE) for a model that predicts risk levels of suffering a traffic accident. Despite the limitations of experimenting with FHE on machine learning models using a low-performance computer, limitations that are undoubtedly overcome by using high-performance computational infrastructure, we built some encrypted models. Among the encrypted models based on Decision Trees, Random Forests, XGBoost, and Fully Connected Neural Networks (FCNN), the model based on FCNN reached the highest accuracy (80.1%) for the lowest inference time (8.476 s).