Improvement of Mixture-of-Experts-Type Model to Construct Dynamic Saliency Maps for Predicting Drivers’ Attention
Sorachi Nakazawa, Yohei Nakada · 2020
Recently, there have been many studies on driving assistance systems to reduce traffic accidents. In such a situation, we proposed a method to construct dynamic saliency maps for predicting driver attention while driving in our previous work. In this construction method, multiple feature maps are first created from the video data of a car-mounted camera. Subsequently, this method computes dynamic saliency maps as combinations of the feature maps that perform center-biasing and normalization. Additionally, this method has been successfully applied to a real-world open dataset, including car-mounted camera videos. However, there is room for further improvement in the model used in this method, since the model cannot capture drivers' characteristics, so that the degree of center bias changes depending on the driving situation. In this paper, we propose an improved model that incorporates multiple center biases into the center-biasing process. It is expected that, by using this improved model, one can appropriately consider changes in the center bias depending on driving situations, and the improved model outperforms the previous model. In addition, the performance of the improved model is compared with that of the conventional model for a real-world open dataset.