Particle Filter Based Correction of On-Board Camera Driver Gaze Estimation

Gaze estimation is an important tool for implementing driver assistance and autonomous driving algorithms. The accuracy of the estimation is crucial considering the impact of failure of such algorithms. In the last decade, an important effort was made to increase this accuracy. However, there are some dynamic conditions that have a great impact on it such as camera vibration determined by road holes or bumps, driver eyes blink or light reflections. The work presented in this paper proposes a modular solution designed to increase the accuracy of gaze prediction independently by the estimation method. It consists of a post-processing step involving a particle filtering solution meant to correct dynamic errors appearing under harsh conditions. To demonstrate the effectiveness of our approach, we use in the validation step an implementation of gaze estimation based on efficient L2CS-Net architecture followed by the particle filter correction module implemented in Python. The test results demonstrate the performance of the presented solution.