Research On Driver Attention Area Extraction And Interpretability Method

Driver attention area extraction is important for dynamic analysis of the driving process and for ensuring traffic safety,and the driver,as a key traffic participant,has a close relationship with road safety in terms of his driving perception and decision-making behaviour.Therefore,we combine vehicle driving data(steering wheel angle,vehicle speed)with the driver’s visual attention area extraction method to carry out research,in order to realize the driver’s perception process based on neural network,driver’s brain perception picture(bird’s eye view vision)simulation,bird’s eye view vision driver’s attention area extraction,real traffic scene driver’s attention area interpretable representation,design experiments and carry out validation,the main work is as follows:(1)Research is carried out on the driver’s attention area extraction method,by perceiving all traffic targets in the driver’s visual area in the real traffic scene image,building a joint traffic scene simulation platform based on the perception results,and realising the virtual bird’s-eye view generation corresponding to the real traffic scene.The fused bird’s-eye view is fitted with the corresponding vehicle driving data for data fitting,the simulation target absence setting is carried out,and the driver’s attention area is extracted in the virtual bird’s-eye view through inversion based on the results of the missing target prediction on the driving data,and the driver’s attention area is visualized on the real traffic scene image using the deep neural network interpretable method.(2)A self-coding network-based lane line detection model and an improved target detection model based on YoloV3 are constructed for the detection of lane lines,vehicles,pedestrians and traffic signals in real traffic scene images.The self-coding lane line detection model,which first addresses the problem of small lane line targets and few pixels,builds a selfcoding network model based on a deep neural network and compresses the image by undersampling.To ensure the accuracy of the segmentation task,the compressed image is recovered through a decoding process to finally achieve lane line detection.The YoloV3 target detection model is improved,and a densely connected deep residual network is designed to improve the feature extraction performance of the model and guarantee the reliability of the target detection results.(3)A joint simulation platform of Apollo 6.0 and LGS VL is built to simulate the real traffic scene environment with LGSVL by combining the lane line detection and target detection results,and a virtual bird’s eye view vision simulation is carried out in Apollo 6.0 based on the LGSVL simulation results to generate a virtual bird’s eye view.A bird’s-eye view visual driver attention area extraction model is designed,and Conv-GRU is introduced to extract temporal features from the data and enhance the extraction of key features with the help of deep neural network attention mechanism.Design the inversion process of driver attention area extraction for bird’s-eye view vision,set different missing targets in the bird’s-eye view vision simulation,use the virtual bird’s-eye view map after the missing targets to re-predict the vehicle driving data,and realise driver attention area extraction in the virtual bird’s-eye view map according to the degree of correlation between driving data and target areas.(4)Designing an attention region extraction model for real traffic scenes and investigating deep neural network interpretable methods.The attention areas in the virtual bird’s-eye view are annotated in the real traffic scene corresponding to the virtual bird’s-eye view,a parallel multi-branch convolutional residual module is constructed for extracting image feature information,and a deep neural network interpretable method is fused to visualise the weighted feature map and to provide an interpretable representation of the driver’s attention areas.The driver attention regions are extracted by expanding the feature map and overlaying it on real traffic scene images through a bilinear interpolation method.