| Pedestrians playing games with passing vehicles when crossing the street is a very common interaction scene in daily traffic.Accurate understanding of pedestrians’ intentions,correct judgment of pedestrians’ crossing trends,and reasonable estimation of the interaction risk between vehicles and pedestrians are not only conducive to perfecting the automobile active safety system,but also helping the driver to perceive the danger in advance and take necessary measures to reduce the possibility of traffic accidents.The hazards of pedestrians can also provide an accurate data basis for the decision-making module and path planning module of autonomous vehicles,further improve the intelligence of autonomous driving,improve ride comfort and safety,promote the advancement of autonomous driving technology.It is of far-reaching significance to ensure the stability of the vehicle,safety and economy.To this end,this research aims to model and estimate the risk of pedestrian-vehicle interaction based on pedestrian multi-modal crossing trajectory prediction.First of all,this paper reviews the current domestic and foreign researches status of pedestrian trajectory prediction and risk modeling of pedestrian-vehicle conflict.It is found that the existing researches on risk modeling seldom consider accurate trajectory prediction,and instead use constant models to roughly estimate the future trajectory or directly predict the probability of collision,and does not consider the multiple trajectory options for pedestrians.Therefore,this paper proposes to predict the multiple possible trajectories of pedestrians by diversifying the possible future motion characteristics of pedestrians,and then based on the time-space relationship of the pedestrian and vehicle trajectories,to realize the modeling of the risk of conflicts that considers more possibilities.The main research contents are as follows:(1)This paper extracts a total of 325 sets of pedestrian-car interaction data in the Euro-PVI dataset and the BPI dataset;Proposes an adaptive absolute coordinate system establishment method,so that the difference in the coordinates of the pedestrian trajectory is mainly derived from its motion characteristics.(2)This paper qualitatively summarizes the characteristics of crossing decision-making and crossing habits;Quantifies the feature vector composed of pedestrian crossing speed,acceleration and heading angle,and obtains 5 kinds of crossing motion states and Gaussian distribution using GMM model;Using LSTM model to predict the possible categories and probabilities of pedestrian motion states in the next0.5s,and the accuracy of the final model for the 5-classification problem is 82.93%.(3)This paper obtains the threshold of the pedestrian’s future motion feature value,and obtains 27 sets of possible future motion feature vectors and occurrence probabilities through uniform resampling;Considering the space-time relationship of people and vehicles,combines the real historical trajectories of people and vehicles with 27 sets of vectors as multiple modal input of the time-space graph transformer pedestrian crossing trajectory prediction model,and obtains multiple possible future trajectories and their probability of occurrence.The ADE of the model is 0.72 m,and the FDE is 0.97 m,which are better than the comparison model.(4)This paper estimates the future trajectory of the vehicle based on the CA model;Selects and calculates the three key indicators including collision probability,minimum encounter distance and minimum encounter distance change rate,establishes a fuzzy mathematical evaluation model,constructs membership functions and divides dangerous membership into five levels: safety,low-risk,medium-risk,high-risk,and extremely high-risk;Human-vehicle interactions in which vehicle has obvious deceleration and avoidance are used for verification.Generally,the hazard can be found 0.5s-1s before the vehicle decelerates.It can more accurately judge the disappearance of danger,which is better than the traditional TTC model. |
PDF Full Text Request