Human-like Evaluation of Visual Perception System for Autonomous Vehicles based on Human Visual Attention

The evaluation of visual perception systems for autonomous vehicles (AVs) is challenging for developing AV technologies. Enlightened by current studies on human-like AV control, assessing the similarity of visual attention between AVs and human drivers becomes one of the potential means for the evaluation of AV technologies. This work proposes a human-like evaluation framework for the visual perception system of AVs, based on human visual attention benchmarks derived from expert human drivers' eye gaze data. We first create a perceptual benchmark by standardizing the gaze data from different expert human drivers. Then, we extract the features from the hidden layers of different AV perception algorithms. Finally, a human-like evaluation index is proposed to quantify the differences between chauffeured vehicles and AVs. The proposed human-like evaluation framework is of help for the development of AV system testing and evaluation by assessing the visual perception of AVs under various driving scenarios.