Research On Saliency Object Estimation Method Based On Driver Attention Prediction

Looking to the future,intelligent connected vehicles are considered as a crucial means to enhance road traffic safety,improve travel efficiency,and promote driving comfort.In order to improve the intelligence level of vehicles,human-like driving is considered to be one of the essential ways.From a bionic perspective,the prediction of driver attention is introduced into the field of automatic driving environment perception.By identifying and locating the driver’s interested objects and areas,potential risks in the driving scene can be quickly and accurately perceived,or key information required for decision-making can be provided,which can enhance the effectiveness and acceptability of the automatic driving system.This article is based on the National Natural Science Foundation of China general project "Research on the method of deformable group perception based on driver scene attention(Project No.52072333)".Aiming at the problem of evaluating the importance of perceived objects in current autonomous driving systems,the article studies driver attention prediction and saliency object estimation in dynamic driving scenarios based on computer vision and human attention mechanism,and proposes a salience object estimation method based on driver attention prediction.1.A driver attention prediction model based on a multiscale spatial-temporal fusion network is proposed.To address the problem of inadequate utilization of spatial-temporal features and the driver attention mechanism in dynamic scenes,a multi-scale spatial-temporal fusion network guided by spatial features is proposed by utilizing attention mechanisms to integrate multiscale spatial-temporal mixed features.The proposed model outperforms other algorithms in multiple saliency metrics on multiple publicly datasets.2.A lightweight driver attention prediction model based on 2D-3D hybrid convolution is proposed.Considering factors such as model size,runtime,and computational resource usage,and the issue of large volume and high computational resource usage of current driver attention prediction models,this paper proposes a lightweight driver attention prediction model through architecture optimization.Through experimental analysis,the proposed lightweight model not only has high prediction performance but also has a model size of 19 MB and a single-frame operation speed of0.02 s,achieving good lightweight and real-time performance.3.A saliency object estimation method based on driver attention prediction is established.Considering the problem that current driver attention prediction is region-level estimation and cannot directly evaluate the saliency of objects,an end-to-end saliency object estimation method is proposed by combining the lightweight driver attention prediction model with object detection algorithms for dynamic driving scenarios.Experimental results show that the proposed method can identify and detect the saliency,category,and position of objects in real-time.