Driver Attention Prediction Targeting To Driving Accident Scenarios

Driver’s attention prediction is an important research topic in human-like driving systems.This thesis attempts to predict the focused area of drivers in driving accident scenarios.However,it is difficult to collect accident scenes online,and the dynamic,complex and unbalanced accident categories also bring great challenges to the task of attention prediction.Therefore,this thesis novelly constructs the driving attention datasets in laboratory,and proposes an attention prediction model algorithm guided by semantic clues of driving scene to fulfill the attention prediction task.(1)Because of the difficulty of the collection of driving accidents data,this thesis collects dashcam driving videos,simulates real driving scenes in laboratory,collects driver eye movement data,and constructs a diverse driving attention benchmark,named as DADA-2000.Each video in this benchmark contains a driving accident,which are divided into 54 categories according to the types of accidents.Totally,we collected 2000 video sequences with 658476 frames of images in the dataset,and the eye movement data are collected for each video frame.This benchmark provides experimental platform for the application of attention prediction in driving scenes.(2)In order to realize attention prediction in complex driving scenes,this thesis designs semantic context induced attentive fusion network(SCAFNet).The network inputs RGB video images and the corresponding semantic segmentation images at the same time.Then,3D convolution network is adopted to extract spatio-temporal features,and the convolution LSTM is used to realize temporal attention transfering.The network can not only characterizes the spatio-temporal characteristics of continuous image frames effectively,but also fuses the scene semantic information of corresponding video frames,i.e.,the spatio-temporal correlation of road participants in driving scenes,thus improving the accuracy of attention prediction in driving scenes.In this thesis,seven indexes,such as KLdiv and CC are used to evaluate the method proposed in this thesis on DADA-2000,DR(eye)VE,and Traffic Gaze datasets,And compared with the seven state-of-the-art attention prediction methods.The results show that the proposed method is more suitable for attention prediction in complex driving scenes.