Human-like Attention-Driven Saliency Object Estimation in Dynamic Driving Scenes

Identifying a notable object and predicting its importance in front of vehicle are crucial for automated systems’ risk assessment decision making. However, current research has rarely exploited the driver’s attentional characteristics. In this study, we propose an attention-driven saliency estimation (SOE) method that uses attention intensity driver as criterion determining salience objects. First, design prediction (DAP) network with 2D-3D mixed convolution encoder–decoder structure. Second, fuse DAP faster R-CNN YOLOv4 at feature level name them SOE-F SOE-Y, respectively, using shared-bottom multi-task learning (MTL) architecture. By transferring spatial features onto time axis, able to eliminate drawback bottom being extracted repeatedly achieve uniform image-video input SOE-Y. Finally, parameters SOE-Y classified into two categories, domain invariant adaptive, then domain-adaptive trained optimized. The experimental results on DADA-2000 dataset demonstrate proposed outperforms state-of-the-art methods several evaluation metrics can more accurately predict attention. addition, driven by human-like mechanism, identify detect salience, category, location objects, providing basis autonomous driving systems.