Högskolan i Skövde

Every year worldwide more than one million people die and a further 50 million people are injured in traffic accidents. Therefore, the development of car safety features that actively support the driver in preventing accidents, is of utmost importance to reduce the number of injuries and fatalities. However, to establish this support it is necessary that the advanced driver assistance system (ADAS) understands the driver’s intended behavior in advance. The growing variety of sensors available for vehicles together with improved computer vision techniques, hence led to increased research directed towards inferring the driver’s intentions. This article reviews 64 driver intention recognition studies with regard to the maneuvers considered, the driving features included, the AI methods utilized, the achieved performance within the presented experiments, and the open challenges identified by the respected researchers. The article provides a high level analysis of the current technology readiness level of driver intention recognition technology to address the challenges to enable reliable driver intention recognition, such as the system architecture, implementation, and the purpose of the technology.

Real-world applications of artificial intelligence that can potentially harm human beings should be able to express uncertainty about the made predictions. Probabilistic deep learning (DL) methods (e.g., variational inference [VI], VI last layer [VI-LL], Monte-Carlo [MC] dropout, stochastic weight averaging - Gaussian [SWA-G], and deep ensembles) can produce a predictive uncertainty but require expensive MC sampling techniques. Therefore, we evaluated if the probabilistic DL methods are uncertain when making incorrect predictions for an open-source driver intention recognition dataset and if a surrogate DL model can reproduce the uncertainty estimates. We found that all probabilistic DL methods are significantly more uncertain when making incorrect predictions at test time, but there are still instances where the models are very certain but completely incorrect. The surrogate DL models trained on the MC dropout and VI uncertainty estimates were capable of reproducing a significantly higher uncertainty estimate when making incorrect predictions.

Driver intention recognition (DIR) studies increasingly rely on deep neural networks. Deep neural networks have achieved top performance for many different tasks. However, apart from image classifications and semantic segmentation for mobile phones, it is not a common practice for components of advanced driver assistance systems to explicitly analyze the complexity and performance of the network’s architecture. Therefore, this paper applies neural architecture search to investigate the effects of the deep neural network architecture on a real-world safety critical application with limited computational capabilities. We explore a pre-defined search space for three deep neural network layer types that are capable to handle sequential data (a long-short term memory, temporal convolution, and a time-series transformer layer), and the influence of different data fusion strategies on the driver intention recognition performance. A set of eight search strategies are evaluated for two driver intention recognition datasets. For the two datasets, we observed that there is no search strategy clearly sampling better deep neural network architectures. However, performing an architecture search improves the model performance compared to the original manually designed networks. Furthermore, we observe no relation between increased model complexity and better driver intention recognition performance. The result indicate that multiple architectures can yield similar performance, regardless of the deep neural network layer type or fusion strategy. However, the optimal complexity, layer type and fusion remain unknown upfront.