| Urban intelligent driving cars inevitably interact with surrounding pedestrians during driving.In this ever-changing environment,pedestrian trajectory prediction and dangerous situation estimation is a key skill to ensure the safe and stable driving of the vehicle.It can provide basic necessary information for vehicle control decisions,greatly improving the scene understanding and interaction capabilities of autonomous vehicles,and enhances the comfort and safety of smart cars.In this paper,pedestrian trajectory prediction and dangerous situation assessment are carried out based on the first perspective of the vehicle.The main research contents are as follows:1.Aiming at the problems of vehicle-mounted pedestrian trajectory prediction,such as the movement of the vehicle-mounted camera,the single feature of pedestrians,and the difficulty of temporal and spatial modeling,a vehicle-mounted pedestrian trajectory prediction algorithm that integrates multiple characteristics and temporal behavior of pedestrians is proposed.By adding richer pedestrian skeleton information,using vehicle global position information and global pedestrian trajectory information,based on the multi-head attention mechanism and channel attention mechanism,the spatial encoder and temporal encoder modules are proposed,which not only extract from time and space dimensions It integrates the temporal and spatial characteristics of the pedestrian’s rich information,and use the time attention module and the space attention module to obtain the pedestrian and self-vehicle,pedestrian and pedestrian latent graph expression structure,and enhance the ability to extract interactive information.In addition,for the fusion of spatio-temporal features,the temporal encoder and spatial encoder are used to further enhance the spatio-temporal features to obtain richer spatio-temporal feature expressions.Finally,the decoder of the transformer network is introduced to further use the predicted trajectory information during decoding.To predict the coordinates of the pedestrian trajectory,thereby enhancing the vehicle’s ability to understand the scene.2.Aiming at the problem of pedestrian dangerous situation estimation,this paper designs a pedestrian dangerous situation estimation algorithm that integrates the uncertainty of pedestrian trajectory.The accuracy of the pedestrian dangerous situation assessment algorithm based on trajectory prediction depends on the accuracy of pedestrian trajectory prediction.However,because the neural network exists in the form of a black box,the trajectory prediction result of the pedestrian trajectory prediction network is uncertain.This uncertainty will lead to the deviation of the system functions of the unmanned vehicle,leading to the safety of the intended function problem.Therefore,this paper introduces the uncertainty evaluation of the network,evaluates the uncertainty of the network through MC-dropout,and improves the pedestrian safety field model based on the uncertainty of the network.Method to complete pedestrian safety posture assessment,thereby enhancing the safety of the intended function of the car.In summary,this article proposes a pedestrian trajectory prediction algorithm that combines multiple characteristics and time-space behaviors of pedestrians in response to the pedestrian safety of smart-driving cars in urban scenes.,and according to the predicted trajectory information,the uncertainty of the predicted trajectory is integrated to complete the dangerous situation assessment of pedestrians.Finally,the effectiveness of this algorithm is verified on the urban smart car driving scene data set and real vehicle experiments. |
