Visual Saliency Regional Prediction Of Traffic Scenes Based On Generative Adversarial Networks

The traffic driving scenes are complex and varied scenes,including important information such as the location of the vehicle and its movement trends,the spatial location of pedestrians and traffic signs.The visual selective attention mechanism is an important neural mechanism for the human visual system to extract important information and filter redundant information in a complex environment.According to this attention mechanism,experienced drivers can efficiently search for and process target information in driving tasks.In recent years,with the research and development of intelligent driving,more and more researchers simulate cognitive activities in driving process through behavioral eye movement experiments,and study the visual attention mechanism and detection model in traffic scenes based on driver eye movement data collected in experiments.In this paper,the eye movement data is used to study the visual information processing characteristics of the driver in the traffic driving scenes,and a traffic scene saliency prediction model based on the generated confrontation network is established.The model can accurately predict the visually significant areas(including the main area and secondary area)of the driver during the simulated driving in the traffic scene.The contents of this paper are divided into the following two parts:In the first part,the source of eye movement data and the preprocessing process of eye movement data are introduced,and the data sets for training and testing of deep learning methods are established.We use basic generation confrontation networks(GAN)and cycle-consistent generative adversarial networks(CycleGAN)to process and analyze the data set,and find that the two basic methods are under-fitting,and the significance evaluation index is also low.Compared with GAN,CycleGAN has better prediction effect,but the design method of generating losses needs to be improved.In the second part,based on CycleGAN’s generation loss and its discriminating method,this paper proposes a progressive training discriminant network model with image scale gradual growth and network multi-step discrimination.This model mainly includes the generation model and the discriminant model,and designs the appropriate generation loss and discrimination methods.The generation model uses a U-shaped mirror structure,including an encoder and a decoder of a progressive structure,in which the network structure of each scale is designed as a residual network unit.The discriminant model performs multi-step discrimination on each scale of the generated image,and gradually corrects the generated image quality.The neural network model proposed in this paper can generate the saliency map of traffic driving scene,which is a significant prediction result.Based on the top-down visual attention mechanism,this paper proposes a predictive model of traffic driving scene saliency based on generative adversarial networks.The model of this thesis can effectively estimate the saliency area of the driver’s visual search and the surrounding environment in the traffic driving scenes.At the same time,it can also point out the main targets of the driver’s gaze and the secondary targets such as traffic signs,which can be used for future intelligence.Provide useful theoretical basis and related technical support for driving vehicles and assisted driving systems.