Fixated Object Detection Based on Saliency Prior in Traffic Scenes

From the perspective of computer vision, both visual saliency and object detection have attracted hot attention in the field of traffic scene perception. However, these two tasks are often seen as independent missions, and their correlations have rarely been explored. In real driving scenarios, drivers mainly care about the salient objects closely related to the current driving task under the guidance of visual selective attention. This process highly integrates saliency perception and object detection, leading to efficient and quick decision-making, thus achieving safe driving. In this study, with reference to human drivers’ perception of traffic scenes, we focus on detecting fixated objects within the regions attracting the drivers’ attention. Firstly, we build a new fixated object detection dataset based on drivers’ fixations, which can serve as a benchmark for studying traffic object detection from the driver’s point of view. Then, we propose a fixated object detection model based on saliency prior, named FOD-Net. FOD-Net takes advantage of the predicted salient regions as saliency priors to guide the detection of the fixated objects that are closely relevant to the driving task, thus improving detection accuracy. 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. Our work combines the driver’s attention mechanism with object detection to narrow the gap between visual saliency and object detection in traffic scenes, showing potential supplemental or referential value for developing high-intelligence assisted/automatic driving systems. The dataset and code are available in <uri>https://github.com/YiShi701/Fixated-object-detection</uri>.