Research On Pedestrian Intention Prediction Method Based On Spatial-temporal Interaction Model

Pedestrian prediction technology is necessary for intelligent and connected vehicles(ICVs),which has been studied extensively in recent years.However,there are still many deficiencies in the existing pedestrian prediction technology,which seriously restricts this technology from applying to ICVs.For example,these researches are constructed by single information,difficult to deal with complex environments and model the spatial-temporal interaction relationship between pedestrians’ historical states.For this reason,this paper takes the pedestrian prediction technology in the ICVs as the research topic,and carries out a pedestrian prediction method based on the spatial-temporal interaction model,which not only can improve the accuracy of pedestrian prediction accuracy,but also can promote the intelligence and safety of ICVs.The main work of this paper is as follows:Firstly,for the problem of the existing pedestrian prediction problem is not clearly defined,a pedestrian prediction framework for ICVs is constructed,the work mainly includes: 1)the functions of pedestrian prediction method for ICVs are analyzed based on the safety and functional requirements;2)taking the function of pedestrian prediction as the starting point,the mathematical model of pedestrian prediction problem is constructed,and the input and output content of pedestrian prediction problem are determined;3)according to the determined input and output information of pedestrian prediction,a pedestrian prediction network model based on deep learning is built.Secondly,the method of pedestrian detection and pose recognition in the complex environment is explored,which includes: 1)the impact of complex lighting on the detection network is studied through qualitative and quantitative analysis methods;2)a multi-illumination image generation method based on generative adversarial network is constructed,which can reduce the impact of complex lighting;3)the generated data are used to train the pedestrian detection and pose estimation model,to increase the accuracy of pedestrian detection and pose recognition in complex environments.Furthermore,a pedestrian intention prediction method based on the spatial-temporal interaction model is constructed,the work mainly includes: 1)an interaction model based on 3D convolution and dense neural network structure is constructed to extract the spatial-temporal interaction relationship of pedestrian historical states;2)the prefusion and post-fusion model are constructed,which can explore the impact of different fusion methods on pedestrian intention prediction;3)the overall training process of the pedestrian intent prediction network is analyzed.Finally,the constructed method and framework are verified,each module is trained by JAAD and COCO datasets,and then verified based on JAAD dataset and real vehicle data.The experimental results on dataset show that each module in this paper has great performance and the proposed prediction framework can effectively predict pedestrian intentions.The experimental results on vehicle data demonstrate the practicability of applying the proposed method to actual scenarios.