Research On Trajectory Prediction Algorithm Of Multi-class Traffic Participants

Vehicle automatic driving technology has become a research hotspot in the automotive industry.Whether the automatic driving vehicle can accurately perceive the surrounding traffic situation and make decisions quickly and reasonably is the key problem faced by the automatic driving technology.The prediction of traffic participant’s trajectory is the intermediate link linking the perception and decisionmaking of the automatic driving vehicle,and the amount of traffic around it based on the perception of the self driving vehicle.It is predicted that the trajectory of traffic participants in the future is the main content of this paper.Urban traffic conditions are often very complex,including small vehicles,large vehicles,non motor vehicles,pedestrians and other traffic participants.At present,it is still a challenging problem to predict the trajectory of multi class participants in urban conditions.In order to solve the problems of low accuracy,inflexible output and low prediction efficiency of multi class traffic participants’ trajectory prediction algorithm,this paper proposes a modular deep learning model,including data preprocessing module,feature extraction module and trajectory prediction module(1)The data pre-processing module,this module includes two parts: trajectory pre-processing and reclassification of traffic participants.The feature of re classification of traffic participants is to use fuzzy clustering algorithm FCM to solve the problem that the original category does not consider the movement characteristics of participants,and divide the participants into three types: radical,stable and conservative.The trajectory pre-processing part summarizes the data processing flow including segment sampling,anomaly recognition and so on;(2)Feature extraction module,which includes two parts: interactive feature extraction and temporal feature extraction,uses graph convolution neural network(GCN)and convolution neural network(CNN)to extract two kinds of features alternately.We improve the construction of object interaction graph of graph convolution,fusing close range interaction graph and in sight interaction graph;(3)Track prediction module,which is three sets of Encoder-decoder recurrent neural network model(seq2seq)for three types of traffic participants,is used to solve the problem of multiple input and multiple output of track data.The author uses the encoder decoder model including LSTM neuron and GRU neuron to compare,and the results show that the prediction efficiency of GRU model is better than LSTM model;(4)This paper uses the Apolloscape trajectory prediction data set for verification and several groups of comparative experiments.The results show that the prediction effect of the proposed model is better than that of the traditional scheme,which improves the accuracy of multi category traffic participants’ trajectory prediction;The problems of the original classification does not consider the motion features,trajectory features are difficult to extract and low efficiency of trajectory output are been solved.