Research On Pedestrian Behaviour Recognition And Collision Warning System Based On Vehicle Vision

As vulnerable road users,pedestrians are a target of particular concern for autonomous vehicles when making decisions.Due to the high degree of mobility and uncertainty of pedestrians,the problem of human-vehicle conflict is particularly acute when pedestrians and autonomous vehicles co-exist in a traffic environment.Further development of autonomous driving systems therefore requires accurate recognition of pedestrian intent in order to reduce the risk of traffic accidents.However,previous research on pedestrian intent has focused on two outcomes,whether a pedestrian crosses the street or not,and lacks consideration of fine-grained behavioural characteristics,which leads to automated driving systems simply making the decision to brake or not to brake,without reflecting the significance of human-vehicle interaction.Therefore,this paper investigates the potential and active interaction behaviour of pedestrians,refines the behavioural features of pedestrians,reveals the connection between the fine-grained behaviour of pedestrians and their actual intentions,so as to achieve pedestrian behaviour recognition,and designs a pedestrian collision warning system based on this,which provides technical support for the safe interaction between autonomous vehicles and pedestrians,and can effectively improve the efficiency of autonomous vehicles on the road.The main work is as follows:Firstly,study the behavioural characteristics of pedestrians in the self-driving perception environment.The implementation scheme of the autonomous driving perception system is determined and data collection from the in-vehicle viewpoint is carried out to realise the analysis of pedestrian behavioural characteristics.Real road sections with high pedestrian traffic and crossing scenarios are selected for the experiments and combined with public datasets on autonomous driving from foreign studies on pedestrian intention for joint statistics.The behavioural characteristics of pedestrians are studied by observing the behavioural patterns and changes of pedestrians in successive frames of video,focusing on the potential interaction behaviour of pedestrians and reflecting their specific intentions through their fine-grained behavioural features.Secondly,as a basis for pedestrian behaviour research,pedestrian detection and tracking needs to be implemented.Based on the study of the YOLOX detection network model,the regression and prediction loss functions are optimised and improved;meanwhile,for pedestrian tracking,a Kalman filtered pedestrian motion prediction model is constructed,and pedestrians in consecutive frames are matched and associated based on the Hungarian matching and BYTE data association algorithms,fusing the improved YOLOX detection model with the tracking model for accurate detection of pedestrians in consecutive frames.After pedestrian detection is achieved,pedestrian intent is identified by fine-grained behavioural features based on the skeleton.A pedestrian interaction behaviour classification system and dataset reflecting the actual intention of pedestrians is constructed,with pedestrian head and hand posture as the main features describing pedestrian behaviour;the key point temporal data of the pedestrian skeleton is extracted through a posture estimation algorithm,and a heat map is stacked along the time dimension to present the human posture features under the image;a 3D residual network and a 3D attention mechanism module are cascaded to extract channel and temporal features and a non-local attention module is introduced into the deep structure of the residual network to enhance the correlation of fine-grained feature variations between successive frames.Finally,a pedestrian collision warning system is designed and constructed based on the previous algorithm study.This specifically includes determining the movement state of the pedestrian by recognising the behaviour of the pedestrian,as well as obtaining parameters such as the distance between the pedestrian and the vehicle.Therefore,firstly,a monocular visual distance measurement model is constructed based on the camera model principle and parameter calibration;secondly,the warning areas of stationary vehicles and moving vehicles are divided respectively;the fuzzy comprehensive evaluation algorithm is used to realise the evaluation of the warning level of pedestrians.Finally,distance measurement and collision warning experiments are carried out to evaluate the pedestrian distance measurement model and the pedestrian collision warning system under different road environments and different driving states of vehicles.The results show that the pedestrians are mainly conservative in crossing the street,and the pedestrians mainly exhibit the implicit interaction behaviour of watching posture while waiting to cross the street,but the pedestrians mainly do not pay attention to the road posture;the improved pedestrian detection and tracking algorithm can improve the pedestrian detection rate and effectively solve the pedestrian occlusion problem;the constructed pedestrian behaviour recognition model is more robust than other methods The self-made pedestrian behaviour dataset has more diverse and complete samples;the distance measurement algorithm is influenced by the road environment and the detection algorithm,and the overall distance measurement accuracy is higher on flat roads;the collision warning algorithm performs relatively well for the warning of vehicles at rest and single-lane scenarios,and the overall accuracy of the warning assessment is higher for the hazard and safety.The early warning algorithm performs relatively well for stationary vehicles and single-lane scenarios,with a high overall accuracy of early warning assessment and a good performance for hazard and safety levels,which to a certain extent can achieve safety protection for pedestrians.