Human Movement Trajectory Completion Technology Based On Deep Learning

Human movement trajectory information is of great significance for the application of location-based services,but affected by various factors,incomplete trajectories may be obtained,which will affect the performance of downstream tasks（such as trajectory matching,etc.）.Therefore,trajectory completion is an urgent problem to be solved.As a relatively new research direction in recent years,there have been a lot of related studies on the trajectory completion problem,but they generally have the following three problems.First,most of the existing models use Recurrent Neural Networks to model the sequential relationship of trajectory points,which makes it difficult to capture the temporal dependencies between various positions and the high-order complex movement patterns in the trajectory,and it is also difficult to detect long-term Dependency modeling.Second,the periodicity of human activities presents a multi-layered feature,which has been ignored in previous work,only modeling simple periodic activities,and not making full use of multi-layered periodicity to complete the trajectory.Finally,the sparse nature of trajectory data brings challenges to the modeling of trajectory data.In order to meet the existing challenges,thesis proposes a human movement trajectory completion model based on graph neural network and attention mechanism for human movement trajectory completion.Specifically,The model includes a position embedding vector representation module,a trajectory relation learning module and a trajectory recovery module.The position embedding vector module,in order to reduce the impact of data sparsity,first constructs a global transition graph between trajectory points based on all trajectories,and uses a graph convolutional neural network to embed the generalized transition relationship between positions into a vector,and then passes Gated graph neural networks model individualized transition relationships within trajectories.The trajectory relationship learning module uses the attention mechanism to learn the relationship within and between trajectories respectively,uses the self-attention mechanism to capture the movement patterns within the trajectory,and uses the cross attention mechanism and soft attention mechanism to capture the multi-level periodicity between trajectories.Finally,the trajectory recovery module fuses all the features to complete the missing positions.In addition,thesis proposes a new loss function that uses spatial approximation for model optimization.In the experimental part,the proposed model is compared with the baseline model on two public datasets in terms of various indicators,and the results verify the effectiveness of the model.And through the ablation experiments to analyze the various parts of the model,it proves the necessity of the existence of each part of the model.