| Pedestrian tracking technology based on video sequence not only has application value in intelligent monitoring,intelligent driving and other fields,but also has high theoretical research significance due to the complexity of pedestrian interaction and the variability of environment.In this theies,from the perspective of pedestrian trajectory prediction,the graph network topology is constructed.Considering the unique time series attributes of video,the data-driven method is applied to integrate the hidden social attributes of pedestrians,so as to improve the prediction accuracy and generalization ability of the pedestrian tracking model.Specific work is as follows:(1)Aiming at the problem of fixed application scenarios and few changes of mathematical models,an unsupervised pedestrian tracking model based on graph convolution is established.RES-LSTM codec architecture is adopted in general.On the one hand,timing characteristics are taken into account,and on the other hand,residual method is introduced to strengthen pedestrian’s current motion intention.First,the graph structure is used to represent the interaction between pedestrians,while considering the direct influence of pedestrians in the field of vision and the indirect influence of pedestrians out of the field,and the influence of pedestrians themselves is set up with the addition of unit matrix,and the adjacency matrix is established to quantify the degree of spatial interaction.Then,the depth graph mutual information algorithm migrated to the graph convolutional network is used as the optimal discrimination algorithm of the model.On the one hand,it maximizes the absorption of individual pedestrian features and on the other hand realizes unsupervised training.The model was verified on ETH and UCY public data sets,and its effectiveness was demonstrated on common ADE and FDE indicators.The average ADE and FDE were the optimal values among the 8 baseline methods,and the multi-modal characteristics of the visualization model were found to predict the trajectory close to the real trajectory.(2)A pedestrian tracking model based on adaptive directed graph network was established to solve the problems of inflexiability in the structure adjustment of dynamic graph,small receptive field in space and time domain,and ignoring the asymmetric interaction between pedestrians in the above model.First,the global feature module is used to broaden the spatial field of perception to break through the limitation of distance threshold under scene changes.Then,the graph convolution plus multi-head mechanism is used to form an attention module to extract the spatial features of pedestrian movement based on distance,direction and speed.Then,the graph attention network is used to adjust the weight of the influence of different historical time tracks on pedestrians’ future tracks.The sequential convolution of row convolution and column convolution maintains the direction and asymmetry,completes the adaptive directed graph learning,splices the spatio-temporal feature results into the global feature selection module to further broaden the temporal field of perception using filling method,and finally,the trajectory tracking module outputs the multi-modal predicted trajectory,considering more scenarios and all of them are close to the real trajectory.It has been proved that the model is effective,and the reasoning time is slightly longer than the above model,but the stability is significantly improved,the fitting speed is faster,ADE and FDE performance is better,ADE and FDE are reduced by 14% and12% compared with the optimal value.(3)Based on the improved model,a pedestrian abnormal behavior recognition algorithm based on trajectory prediction is proposed,which is applied to the identification and detection of crossing the warning line,entering the danger zone,staying and gathering,and the design of pedestrian monitoring electronic fence is completed.The system meets the requirement of real-time and stability,and the recognition accuracy is more than 95%. |
