Pedestrian Intention Estimation And Trajectory Prediction Driven By Data And Knowledge

In recent years,autonomous driving technology is in a period of rapid development.It is foreseeable that in a period of time in the future,autonomous driving vehicles will inevitably share traffic roads with pedestrians.Pedestrians are a vulnerable group of road participants,and their safety is currently the biggest challenge in the application of autonomous driving technology.However,an accurate and efficient pedestrian intention estimation and trajectory prediction technology is the key to solving the pedestrian safety problem,which has important research significance for autonomous vehicles.At present,most of the research in this area adopts a purely data-driven method,but it requires a large number of data samples,has the characteristics of a black box,and cannot understand pedestrian intentions and trajectories in the way humans perceive pedestrian crossing scenes.Therefore,this paper proposes a data-plus-knowledge-driven pedestrian intention estimation and trajectory prediction method,using the pedestrian crossing scene text corpus as the knowledge resource,and using related technologies to complete the construction of the pedestrian crossing scene knowledge map,based on the built map combined with external reasoning technology and scene Semantic similarity matching technology establishes a reasoning mechanism for pedestrian crossing intention estimation,and integrates pedestrian intention information into a multi-feature trajectory prediction model to complete the prediction of pedestrians’ future trajectory.The main research work is as follows:(1)Construction of knowledge graph for pedestrian crossing scene.Aiming at the problem that there is almost no public knowledge map of pedestrian crossing scenes,this paper firstly uses the description information in the PSI dataset and JAAD dataset as the basis,combined with the description text in the relevant literature in the field of autonomous driving,and completes the corpus with 7372 pieces of corpus after preprocessing.Describe the establishment of the text library,use the Doccano platform to complete the text annotation,and build a targeted triplet dataset of 10 entity data categories and 58 relational data categories.Secondly,an improved CASREL triplet extraction model is proposed.The experimental results show that the extraction effect of the F1 index is 89.4%,which is 3.0%higher than the original model.It can effectively solve the problem of entity overlap and complete the pedestrian crossing scene.Triple extraction task describing text.Finally,construct a pedestrian crossing scene ontology model with 8 major categories and 26 small categories of entity concepts.After fusing the abstract relationship and commonsense information between entity concepts,a model layer is formed.Through the knowledge processing and quality of triplet collection Evaluation,high-quality triplet information is obtained,and an online knowledge graph of real-time scenes and an offline knowledge graph with 736 triplets and 20 pedestrian crossing sub-scene graphs are constructed.(2)Estimation of pedestrian crossing intention based on scene semantic similarity matching and KG＿BN.Aiming at the problems of complex models and low interpretability in pure datadriven methods in pedestrian crossing intention estimation research,this paper proposes a knowledge graph(KG)mapping Bayesian Network(BN)data.The driving reasoning method is the main method,and the knowledge-driven analogy reasoning method of scene map semantic similarity matching is supplemented by the reasoning mechanism of pedestrian crossing intention estimation.For the BN model,through structure and parameter learning,nine characteristic variables that affect pedestrian crossing intention are determined,and their probability distributions are obtained to complete the construction of the BN model.The experimental evaluation results show that the F1 value of the model is 88.76%,and the AUC value is 0.94,which has a good effect on the way of directly estimating the intention of pedestrians to cross the street with the data-driven method;Whether pedestrians cross the street has a significant impact.For each sub-scene graph in the offline knowledge graph of the pedestrian crossing scene,the BN model is used to integrate the probability of pedestrian crossing into it respectively,and an offline knowledge base with probability is obtained.On this basis,aiming at the scene map semantic similarity matching mechanism that simulates the human analogy reasoning method,a GraphSAGE graph neural network model that extracts the global features of the scene map is constructed.The experimental results show that when the similarity threshold is 80%,the AUC value can reach 0.88,which has a good similarity classification effect,combined with examples to verify the effectiveness of the knowledge-driven analogy reasoning method in indirectly estimating pedestrian crossing intentions.Compared with the purely data-driven method,the pedestrian intention estimation mechanism constructed in this paper can simulate human perception of pedestrian crossing scenes,and has better interpretability and accuracy.(3)Trajectory prediction that integrates pedestrian behavior intention information.Aiming at the problem that the pedestrian trajectory prediction model considers the single feature information and the sudden change of pedestrian behavior intention affects the trajectory,which leads to the poor prediction effect of the model,this paper proposes a pedestrian trajectory prediction model based on multi-feature information that combines pedestrian behavior intention information knowledge inference and BiLSTM+Attention data-driven method.Aiming at the behavior intention of pedestrians,an inference rule base is established.After the semantic transformation of the scene map information,the SWI-Prolog inference engine is used to complete the knowledge reasoning of pedestrian turning behavior intention information and detour behavior intention information,and the effectiveness of the reasoning method is verified by examples.Aiming at the experimental requirements of pedestrian trajectory prediction,build a virtual street crossing scene,use VR equipment to complete data collection,and construct the trajectory data set of the virtual scene after relevant processing.The comparison experiment of different prediction time shows that the prediction effect of the first 50%of the observed sequence data in this paper is the best after 50%of the data.The ADE index of the model experiment is 0.376,and the FDE index is 0.689.The comparative experiments of different characteristic factors show that,on the basis of trajectory features and head posture angle information,after incorporating information such as pedestrian behavior intentions,the prediction accuracy of the model is significantly improved.The ADE index of the model experiment decreased by 20.8%,and the FDE index decreased by 25.4%,compared with the trajectory prediction model that considers single feature information,the pedestrian trajectory prediction model that integrates multiple features constructed in this paper has better accuracy.