Improving Computational Method of Dynamic Saliency Maps for Estimating Drivers’ Fixation Points Based on Issues Extracted via In-vehicle Camera Video Simulator

Currently, driving support technologies are making notable progress in reducing road accidents caused by drivers’ errors or inattention. To further develop the driving support technologies, we have been continuously working on a statistical-machine-learning-based computational method of dynamic saliency maps for estimating drivers’ fixation points, which enables us to quantify the degree of drivers focusing on specific objects as a probability. Simultaneously, we have also developed a prototype simulator to replicate in-vehicle camera videos under a variety of driving situations using 3-dimensional computer graphics to validate such a computational method of dynamic saliency maps. Moreover, in our previous study, we have extracted two issues of the above-mentioned computational method via validations with in-vehicle camera videos replicated using this simulator. In the present study, we improve the computational method by introducing techniques for overcoming these extracted issues. The computational method improved using these techniques is also tested with real videos captured using real in-vehicle cameras and replicated videos using our simulator. Based on these results, we discuss the effectiveness of the improved method.