Research On Pedestrian Intention And Trajectories Prediction Based On Forward Looking Scenes Of Vehicles

With the increase of car ownership,the traffic safety problem is becoming more and more serious.Intelligent vehicles can improve driving safety through reasonable and effective decision-making methods,thereby reducing the occurrence of traffic accidents.As an important traffic participant,pedestrians’ intention and trajectory will inevitably affect the decision-making and planning of intelligent vehicles.Therefore,research on pedestrian intention and trajectory prediction is of great significance to improve the driving safety of intelligent vehicles.Aiming at the problem that pedestrians under the front view of the vehicle have variable motions,this paper conducts in-depth research on pedestrian intention and trajectory prediction algorithms.The main contents are as follows:(1)A pedestrian intention prediction framework by fusion multi-features based on machine learning is constructed: Firstly,pedestrian skeleton information is extracted through a top-down human pose estimation algorithm to describe the motion of the pedestrian.Then,head orientation estimation algorithm is used to obtain head direction information,and head direction is fused with skeleton features to enhance the expression of pedestrian motion characteristics.Finally,the intention prediction result of the fusion of multiple features is obtained by the long short-term memory(LSTM)network.The experimental results show that the fusion multi-feature intention prediction method has an accuracy rate up to 96.0%,which is better than the single feature intention prediction network.In the actual scene analysis,the proposed method can identify the pedestrian’s bending intention 0.56 s in advance,and can provide sufficient redundant time for decision-making of intelligent vehicles.(2)A hierarchical pedestrian trajectory prediction framework fused with pedestrian intentions is proposed: To reduce the impact of pedestrian movement uncertainty on trajectory prediction performance,an I-LSTM trajectory prediction network fused with pedestrian intentions is constructed in this study.The fusion vector of the prediction result obtained by the intention prediction layer and the historical trajectory coordinates is used as the input of the trajectory prediction layer to generate the trajectory prediction result.In addition,the attention mechanism is introduced in the trajectory prediction network to enhance the effective use of the LSTM encoding vector at each time and improve the performance of pedestrian trajectory prediction.The experimental results show that the RMSE of the position regarding the trajectory prediction is 347 mm with a prediction horizon of 1 second,which is better than that of baseline methods such as constant velocity model(CV),interacting multiple model(IMM),conventional LSTM and generative adversarial networks(GAN).In the actual scene analysis,the method proposed can significantly reduce the trajectory prediction error in the process of pedestrian intention change,and make pedestrian participating in traffic safer.(3)A variety of pedestrian intention fusion schemes are proposed: This paper proposes three intention fusion schemes,including pre-intention fusion,post-intention fusion and intention as a model selection mechanism.The pre-intention fusion scheme fuses the intention and historical trajectory at the input of the trajectory encoder.The post-intention fusion scheme fuses the intention features at the output of the encoder.The intention as a model selection mechanism scheme needs to train multiple trajectory prediction models,and then select the corresponding model according to the intention to achieve targeted trajectory prediction.Experimental results show that three intention fusion methods proposed can effectively fuse the intention features and improve the accuracy of trajectory prediction.Among them,the pre-intention fusion scheme can more effectively cope with the needs of intent change scenarios,and its robustness is better.