Abstract: PURPOSE: Driving safety is the most important for assisted/autonomous driving. Referring to the driver’s perception of the traffic scene, we combine the driver’s selective attention mechanism with computer vision to improve the detection performance of the vital fixated objects closely related to the driving task, providing the potential application or referential value in the intelligent driving safety.
METHODS: Based on the fixations of more than 28 experienced drivers, we first build a new eye-tracking-based fixated object detection dataset (ETFOD). Then, we propose a fixated object detection model based on saliency prior, named FOD-Net. It consists of three parts: the object detection module (ODM), the salient region guided module (SRGM) and the saliency guidance strategy. ODM is a strong baseline detector responsible for detecting traffic fixated objects with various scales. SRGM predicts pixel-wise saliency maps in shallow layers, which contain detailed salient regions attracting drivers’ attention. Finally, based on the saliency guidance strategy, the salient regions generated by SRGM can be used as saliency priors to guide ODM to pay more attention to the fixated objects within the salient regions, thus enhancing ODM’s feature representation for fixated objects instead of objects outside the drivers’ attention regions. The two tasks of fixated object detection and salient region prediction can mutually facilitate each other to improve fixated object detection accuracy.
RESULTS: Experimental results on the proposed dataset show that FOD-Net achieves a mAP value of 78.4% with small model parameters, which is higher than other state-of-the-art models.
CONCLUSIONS: Combining the driver’s attention mechanism and the object detection together can achieve more accurate detection for the fixated objects with direct threats to driving safety, showing potential application value for developing high-intelligence assisted/automatic driving systems.

Key words: Traffic object detection, Selective attention, Object detection dataset, Deep learning.