A Hybrid Approach for Turning Intention Prediction Based on Time Series Forecasting and Deep Learning

Kernelized convolutional transformer network based driver behavior estimation for conflict resolution at unsignalized roundabout

The modeling of driver behavior plays an essential role in developing Advanced Driver Assistance Systems (ADAS) to support the driver in various complex driving scenarios. The behavior estimation of surrounding vehicles is crucial for an autonomous vehicle to safely navigate through an unsignalized intersection. This work proposes a novel kernelized convolutional transformer network (KCTN) with multi-head attention (MHA) mechanism to estimate driver behavior at a challenging unsignalized three-way roundabout. More emphasis has been placed on creating convolution in non-linear space by introducing a kervolution operation into the proposed network. It generalizes convolution, improves model capacity, and captures higher-order feature interactions by using Gaussian kernel function. The proposed model is validated using the real-world ACFR dataset, where it outperforms current state-of-the-art in terms of behavior prediction accuracy and provides a significant lead time before potential conflict situations.

Driver's turning intent recognition model based on brain activation and contextual information

Traffic situations like turning at intersections are destined for safety-critical situations and accidents. Human errors are one of the main reasons for accidents in these situations. A model that recognizes the driver's turning intent could help to reduce accidents by warning the driver or stopping the vehicle before a dangerous turning maneuver. Most models that aim at predicting the probability of a driver's turning intent use only contextual information, such as gap size or waiting time. The objective of this study is to investigate whether the combination of context information and brain activation measurements enhances the recognition of turning intent. We conducted a driving simulator study while simultaneously measuring brain activation using high-density fNIRS. A neural network model for turning intent recognition was trained on the fNIRS and contextual data. The input variables were analyzed using SHAP (SHapley Additive exPlanations) feature importance analysis to show the positive effect of the inclusion of brain activation data. Both the model's evaluation and the feature importance analysis suggest that the combination of context information and brain activation leads to an improved turning intent recognition. The fNIRS results showed increased brain activation differences during the "turn" decision-making phase before turning execution in parts of the left motor cortices, such as the primary motor cortex (PMC; putative BA 4), premotor area (PMA; putative BA 6), and supplementary motor area (SMA; putative BA 8). Furthermore, we also observed increased activation differences in the left prefrontal areas, potentially in the left middle frontal gyrus (putative BA 9), which has been associated with the control of executive functions, such as decision-making and action planning. We hypothesize that brain activation measurements could be a more direct indicator with potentially high specificity for the turning behavior and thus help to increase the recognition model's performance.

Three Layered Architecture for Driver Behavior Analysis and Personalized Assistance with Alert Message Dissemination in 5G Envisioned Fog-IoCV

The Internet of connected vehicles (IoCV) has made people more comfortable and safer while driving vehicles. This technology has made it possible to reduce road casualties; however, increased traffic and uncertainties in environments seem to be limitations to improving the safety of environments. In this paper, driver behavior is analyzed to provide personalized assistance and to alert surrounding vehicles in case of emergencies. The processes involved in this research are as follows. (i) Initially, the vehicles in an environment are clustered to reduce the complexity in analyzing a large number of vehicles. Multi-criterion-based hierarchical correlation clustering (MCB-HCC) is performed to dynamically cluster vehicles. Vehicular motion is detected by edge-assisted road side units (E-RSUs) by using an attention-based residual neural network (AttResNet). (ii) Driver behavior is analyzed based on the physiological parameters of drivers, vehicle on-board parameters, and environmental parameters, and driver behavior is classified into different classes by implementing a refined asynchronous advantage actor critic (RA3C) algorithm for assistance generation. (iii) If the driver’s current state is found to be an emergency state, an alert message is disseminated to the surrounding vehicles in that area and to the neighboring areas based on traffic flow by using jelly fish search optimization (JSO). If a neighboring area does not have a fog node, a virtual fog node is deployed by executing a constraint-based quantum entropy function to disseminate alert messages at ultra-low latency. (iv) Personalized assistance is provided to the driver based on behavior analysis to assist the driver by using a multi-attribute utility model, thereby preventing road accidents. The proposed driver behavior analysis and personalized assistance model are experimented on with the Network Simulator 3.26 tool, and performance was evaluated in terms of prediction error, number of alerts, number of risk maneuvers, accuracy, latency, energy consumption, false alarm rate, safety score, and alert-message dissemination efficiency.