Research On Driver’s Fixated Object Detection Methods Based On Visual Selective Attention

Traffic scene perception refers to the awareness and understanding of the surrounding traffic environment by road participants,including the identification of objects on the road and the interaction between these objects.Inexperienced or distracted drivers often fail to allocate attention properly to important traffic elements,leading to deviations in their perception and understanding of key information during the driving process,thereby increasing the risk of traffic accidents.In contrast,experienced drivers are influenced by top-down factors(e.g.,driving task,traffic rules,accumulated experience,etc.)and bottom-up factors(e.g.,motion,color,etc.).They focus on objects closely related to the driving task and filter out irrelevant information to make correct decisions in a timely manner,thus achieving safe driving.Unlike existing computer vision methods that treat visual saliency and object detection as separate tasks,the human visual system effectively accomplishes these tasks by sharing information processing across various visual tasks.Referring to drivers’ perception of traffic scene and combining research findings in computer vision,this dissertation focuses on detecting objects that attract drivers’ attention,referred to as fixated object detection.To conduct this research,the eye movement characteristics and attention distribution of drivers are firstly delved,and then static image and dynamic video datasets suitable for traffic fixated object detection are established.Finally,targeted methods are proposed from the perspective of computer vision to achieve effective detection of fixated objects.The main research contents and contributions of this dissertation are as follows:(1)The allocation of visual attention by drivers during the driving process is investigated.By analyzing the eye movement signals of multiple experienced drivers during simulated driving tasks,it can be observed that under the modulation of top-down attentional mechanism,drivers are able to adaptively focus their attention on important traffic elements according to changes in traffic scenes and driving behaviors.Simultaneously,affected by the bottom-up attention mechanism,drivers can promptly attend to unexpected factors in the driving process,such as overtaking from the side.Based on the eye movement experiments in cognitive psychology,eye movement data of multiple drivers collected in real driving tasks is analyzed in detail.The results indicate that drivers exhibit similar eye movement distribution characteristics and patterns under the guidance of visual selective attention mechanism.The study on drivers’ eye movement characteristics and attention distribution provides important references for conducting fixated object detection.(2)This dissertation conduct research on fixated object detection in static traffic images.Based on the eye movement information of drivers,a new fixated object detection dataset is established,which can serve as a benchmark for studying traffic object detection from the perspective of drivers.Subsequently,by referencing the drivers’ perception of traffic scenes and combining with the research findings on computer vision,a fixated object detection model is designed.The proposed model decomposes fixated object detection into two subtasks: visual saliency prediction and object detection.The salient regions that attract the driver’s attention are predicted by sharing intrinsic correlated information required for multiple tasks.The predicted regions can serve as saliency priors to guide the model to focus on traffic fixated objects selectively,thereby enhancing the model’s feature representation of fixated objects.Experimental results demonstrate the effectiveness of the proposed method,and the constructed model achieves more competitive detection performance compared to other state-of-the-art models.(3)The research on fixated object detection in dynamic traffic videos is conducted.Based on eye movement data of drivers,a new dataset for fixated object detection in traffic videos is established,serving as a benchmark for video fixated object detection.Inspired by drivers’ visual selective attention mechanism,a fixated object detection model for traffic videos is constructed.The shallow saliency and deep global information closely related to driving tasks are decoded from reference frames.The decoded saliency and global information serve as top-down priors to enhance the model’s encoding weights for fixated objects in the current frame.The experimental results indicate that,compared with other state-of-the-art object detection models,the proposed model achieves higher detection accuracy while maintaining real-time detection speed.Considering the high cost of manually labeled data,this dissertation further develops a weakly supervised method for fixated object detection in traffic videos.Experimental results manifest that the proposed weakly supervised method achieves detection performance close to that of some fully supervised models.(4)To address the challenge brought by the significant scale variation of fixated objects to the detection model,this dissertation proposes a simple yet effective detection head configuration method to represent the categories and positions of fixated objects.In summary,this dissertation conducts research on fixated object detection around the perception and understanding of traffic scenes.The established datasets provide benchmarks for attention-inspired traffic fixated object detection.The proposed methods enhances the detection ability of fixated objects closely related to driving tasks.The competitive detection performance demonstrates potential application value in improving the safety of intelligent driving or assisted driving,providing new ideas for the development Existing detection head configuration methods are typically based on default values or personal experience.When input resolution or detection tasks change,these methods cannot efficiently and quantitatively configure detection heads,resulting in the degradation of detection performance.The appropriate detection heads are selected to optimize existing detection models according to the matching degree between the object scale and potential detection heads,thereby effectively improving the detection accuracy of the models.Considering the complexity of the traffic scene and the limitation of on-board computing resources,a skip-scale detection head configuration guideline is developed.The results indicate that the proposed method can use a small number of detection heads instead of multiple ones,reducing model parameters while maintaining high detection accuracy and improving inference speed.of advanced intelligent or assisted driving systems.