Representing Realistic Human Driver Behaviors using a Finite Size Gaussian Process Kernel Bank

Abstract

The performance of cooperative vehicular applications is tightly dependent on the reliability of the underneath Vehicle-to-Everything (V2X) communication technology. V2X standards, such as Dedicated Short-Range Communications (DSRC) and Cellular-V2X (C-V2X), which are passing their research phase before being mandated in the US, are supposed to serve as reliable circulatory systems for the timecritical information in vehicular networks; however, they are still heavily suffering from scalability issues in real traffic scenarios. The technology-agnostic notion of Model-Based Communications (MBC) has been proposed in our previous works as a promising paradigm to address the scalability issue and its performance, while acquiring different modeling strategies, has been vastly studied. In this work, the modeling capabilities of a powerful nonparametric Bayesian inference scheme, i.e., Gaussian Processes (GPs), is investigated within the MBC context with more details. Our observations reveal an important potential strength of GPbased MBC scheme, i.e., its capability of accurately modeling different driving behavioral patterns by utilizing only a limited size GP kernel bank. This interesting aspect of integrating GP inference with MBC framework, which has been verified in this work using realistic driving data sets, introduces this architecture as a strong and appealing candidate to address the scalability challenge. The results confirm that our proposed approach overperforms the state of the art research in terms of the required communication rate and GP kernel bank size.Index Terms-Vehicular ad-hoc network, scalable V2X communication, model-based communication, non-parametric Bayesian inference, Gaussian processes, driver behavior modeling.

Context‐aware target classification with hybrid Gaussian process prediction for cooperative vehicle safety systems

“…Additionally, we want to mention that our approach can benefit from parallel computing during the bank kennel creation and during forecasting, as the model predictions can be computed independently in parallel, therefore the use of GPU or a CPU with additional cores will further reduce the computation time. Based on our experiments and previous work [51], we choose a $$TW$$ size of 30 time‐steps (3 s) and $$PTE_{th}$$ = 50 cm for the training, model generation, and forecasting. The selected training parameters consider an acceptable path history while also meeting the computational requirement of the safety applications.…”

Robustness and Adaptability of Reinforcement Learning-Based Cooperative Autonomous Driving in Mixed-Autonomy Traffic

“…Social navigation in mixed autonomy has shown the potential of collaboration among AVs and HVs [27]. Current works in social navigation tackle the MARL cooperation by assuming the nature of agent interactions [28], [29] or by directly modeling or classifying human driver behaviors [30]- [32]. Different methods to predict or classify driver behaviors are based on driver attributes [33], graph theory [34], game theory [1] and data mining [35].…”

Robustness and Adaptability of Reinforcement Learning based Cooperative Autonomous Driving in Mixed-autonomy Traffic

“…Social navigation in mixed autonomy has shown the potential of collaboration among AVs and HVs [40]. Current works in social navigation tackle the MARL cooperation by assuming the nature of agent interactions [41], [42] or by directly modeling or classifying human driver behaviors [43]- [45]. Different methods to predict or classify driver behaviors are based on driver attributes [46], graph theory [47], game theory [1] and data mining [48].…”