A Predicting Human Eye Fixations Model For Task-oriented Routing In Transportation Hubs

In a transportation hub,people often need to use relevant information in visual scenes,such as building space,signs,crowd directions and landmarks,when they are looking for the way with related tasks such as finding the gate or gate.Human visual attention mechanism enables people to quickly and accurately screen out important information related to the current task when processing complex visual scenes.The brain prioritizes processing and analysis of this visual information to complete the visual search task.At this time,the quality of architectural design is particularly important.However,there are great limitations in evaluating the quality of architectural design through methods such as field survey and eye movement tracking.The significance model can easily and repeatedly carry out the pre-evaluation of pathfinding quality under the same type of architecture for multiple projects.Most of the existing significance detection models are based on data-driven,bottom-up attention mechanisms.However,the prediction of gaze points in the traffic junction scenario we studied is based on the top-down attention mechanism of road finding behavior.People usually carry out the search task with the cognition of the scene context and the prior knowledge of the task,such as the shape,color and possible position of the target,which can help people quickly direct their visual attention to the task target.However,most of the existing eye-movement data sets were collected when the subjects were free to watch the natural scene and fixed the viewing time.Therefore,they could not extract the prior information needed for this subject.Therefore,for the subject of this study,we need to design eye movement experiment,collect an eye movement data set based on the pathfinding task in the scene of traffic junction,analyze the visual behavior of people in the process of pathfinding,and obtain prior information.Then,by combining the low-level significance,scene context and task prior features,a prediction model of gaze points driven by pathfinding task in traffic junction scene is constructed.The main research contents of this paper are as follows:(1)The eye movement data set of traffic hub pathfinding task was established.First,the indoor scene images of transportation hubs were collected.Then,a pathfinding task-oriented eye movement experiment was designed.Eye movement data was collected by eye tracker to generate the required documents for the data set.In addition,we studied the eye movement characteristics of people in this scene task,extracted thetranscendental information related to the target--vanishing point,and proved the transcendental guiding effect of vanishing point on this task.(2)A vanishing point detection method VOD based on weber local descriptor was proposed.Aiming at the image characteristics in the scene of traffic hub,the weber local descriptor is introduced and improved.The positive and negative difference excitation is separated to retain the change characteristics of the gray level,and Gabor filter is used to detect the gradient direction,so as to fully reflect the spatial distribution information in the local window,effectively distinguish the interference and clues,and obtain the effective voting area.Then a linear voting model with prior constraints is used to obtain the position of vanishing point.In this paper,an experiment is conducted on the data set of vanishing point under the traffic junction scene marked by the author.The experimental results show that,compared with the existing detection methods of vanishing point,the detection method in this paper is more suitable for the traffic junction scene.(3)The prediction model of eye fixation point based on the task of finding the way of traffic hub is established.The model combines low-level significance,scene context and task prior features,including three parts: prediction module of initial fixation,auxiliary module of prior information and module of prior information fusion.The initial fixation prediction module is responsible for extracting the low-level significance and context information to obtain the initial fixation prediction graph.At the same time,the prior information auxiliary module is responsible for extracting the vanishing points in the image,and the gaussian weight is introduced to generate the prior map of the vanishing points.Finally,through the prior information fusion module,the obtained prior features are weighted and fused with the prediction graph of the initial fixation point,so that the final significance graph is closer to the significant truth value.We compared the model in this paper with other 7 models,and the experiment on the eye movement data set under the traffic hub pathfinding task shows that the model in this paper is more consistent with the human gaze behavior than the existing gaze point prediction model.