Research On Representation Learning And Prediction Model Of Crowd Movement Trajectory Based On Deep Learning

The trajectory of people's movement reflects people's travel rules,and trajectory representation learning and prediction is one of the important research directions in the field of spatio-temporal data mining.It is important for understanding human behavior patterns,studying unmanned technology,urban planning,alleviating urban traffic congestion and the key population management.In recent years,with the development of artificial intelligence,many trajectory modeling methods use deep learning models to extract feature representations of inherent pattern from complex movement trajectory data to improve the accuracy of trajectory prediction.At present,deep learning has achieved certain results in the fields of traffic flow prediction,traffic jam recognition,and destination prediction.However,most of the trajectory modeling methods based on deep learning are directly transferred from the fields of natural language processing or image processing,and cannot robustly represent the inherent pattern of trajectory data.On the one hand,the trajectory data contains the internal features of the coordinates and the contextual features of the trajectory,and the unified representation method of the two parts of the features needs to be considered.On the other hand,trajectory data has the features of long-sequence spatio-temporal interaction,and the process of spatiotemporal evolution needs to be considered.Therefore,studying the fusion representation method of the spatio-temporal interaction features and the internal pattern of the trajectory can not only improve the prediction ability of the deep learning model,but also help to explore the pattern of the trajectory in geographic space,and provide a scientific method for exploring the pattern interaction process of the trajectory.This paper takes the crowd movement trajectory collected by the base station as the research object.The composition of the spatio-temporal feature of trajectory is studied,including coordinate sequence,crowd interaction,and remote dependence at different scales.Then,the trajectory prediction model based on deep learning is constructed.The main contents of this paper are as follows.(1)Aiming at the problem that coordinates and trajectory context features cannot be represented jointly in real space,a Trajectory Representation Learning Method in Complex Space Based on Skip-gram(TRMC+Skipgram)is proposed.First,the trajectory context relationship is transformed into a graph structure to solve the problem of trajectory context discontinuity.Second,based on Euler's formula,a complex vector distance calculation method is derived,combined with Skip-gram and graph sampling methods,to learn the coordinate and trajectory context feature joint representation vector in the complex space.Finally,in the node linking prediction task,TRMC+Skipgram can effectively aggregate coordinates with similar features.In the trajectory-user linking task,it can also effectively improve the accuracy of the trajectory-user linking.(2)Aiming at the problem of trajectory missing caused by irregular sampling of base stations,combining two standard Long Short-term Memory(LSTM),a LSTM in Complex Space(CLSTM)is proposed.A Missing Trajectory Data Inference Model Based on CLSTM(MTDIC)is constructed.First,MDIC loads the coordinate pre-training vector learned by TRMC+Skipgram in the coordinate embedding layer,and incorporates the long-short period features of the trajectory.The interaction features of the trajectory pre-training vector are extracted through CLSTM,and the non-linear mapping layer is combined to realize the completion of the missing trajectory.Second,based on the nonlinear mapping of the low-dimensional potential subspace and the schatten norm,a kernel regularization loss function is derived to ensure that the trajectory data recovered by MDIC meets the low rank.Finally,in a year's trajectory data of 100,000 people in a city and a public dataset Gowalla,MDIC is verified,which can learn the movement patterns of the crowd and improve the accuracy of the restoration trajectory.(3)Aiming at the problem that the traditional trajectory destination prediction model does not consider the interaction features of the crowd,a Spatio-temporal Attention Deep Learning Prediction Model(STAM)is proposed.The model includes a temporal attention module,a spatial attention module,a spatio-temporal feature fusion module,and a destination prediction module.First,a temporal attention module is designed based on multi-head attention to extract the temporal interaction features of the trajectory.Second,a spatial attention module is designed based on the graph attention neural network to extract the spatial interaction features of the crowd.Then a spatio-temporal feature fusion module is constructed to adaptively control the fusion of temporal and spatial interaction features.Finally,the trajectory destination prediction module maps the spatio-temporal interaction features to the trajectory destination.STAM is verified on the two public datasets of ETH and UCY,which can effectively extract the temporal interaction features of the trajectory context and the spatial interaction features of the crowd,and reduce the error between the real trajectory destination and the predicted trajectory destination.(4)Aiming at the high exposures bias caused by the multi-step prediction model of the training trajectory of teacher forcing,a nested iterative decoding algorithm is proposed,which is combined with the Sequence to Sequence(Seq2Seq)framework to construct a Trajectory Multi-step Prediction Model Based on Nested Iteration(TPNI).First,TPNI uses parallel temporal features extraction module and spatial features extraction module to extract lowcoupling spatio-temporal interaction features.Then,it uses multiple decoders to generate multiple tracks with different time starting points by nesting decoders on the time axis to realize the updating of its parameters.Finally,TPNI is verified on the dataset recovered by MTDIC,which can effectively reduce the error of multi-step trajectory prediction.(5)The research method is applied to the smart city personnel management platform.A crowd heat map module is constructed based on complete crowd trajectory data recovered by TRMC+Skipgram and MTDIC for regional population density analysis.At the same time,in order to manage high-density areas of personnel,a rail transit pedestrian flow management module is constructed based on STAM to predict the getting off stations of people taking rail transit,and combined with crowd heat map information to realize overload warning at each station.Finally,in order to monitor the behavior trajectory of key management crowd,a key crowd management module is constructed based on TPNI,combined with the crowd heat map module and the rail transit pedestrian flow management module provides high density area information,to early warning of abnormal behavior of the crowd,realizes the intelligent wisdom city personnel management.