Trajectory Prediction Based On Deep Learning

With the popularity of smartphones and GPS navigators,massive data which contains location information is constantly being generated.Therefore,the need for real-time analysis and full utilization of these location data is increasingly urgent.As the most common location information data,trajectory data is the focus of research in the Intelligent Transportation System(ITS).Since real-time prediction of the trajectory of moving objects can effectively prevent traffic congestion and provide users with more personalized location-based services.There are solutions for various trajectory prediction problems,including conventional statistical-based methods and deep learning-based methods.However,existing methods lack a unified evaluation environment.At the same time,many algorithms use external data,which is often difficult to obtain,to improve the prediction accuracy.It leads to a lack of fair experimental settings and evaluation criteria when comparing these algorithms.For these problems,we conduct a comprehensive evaluation of various mainstream prediction algorithms in the same dataset,hardware environment,experimental settings and evaluation criteria.We selected 5 statistical method based algorithms and 10 deep learning-based algorithms.The evaluation background and setting we give is to use the trajectory data of a moving object's past 30 timestamps to predict the future trajectories for different lengths and directly use raw trajectory data without any external data.Using average absolute error and hit rate to uniformly evaluate the prediction effect.We use three real trajectory datasets with different motion patterns to conduct extensive empirical evaluations of the above 15 algorithms and reveal interesting findings and patterns based on experimental results.When evaluating each trajectory prediction algorithm,we found two problems.One is that there are a large number of static trajectories that match the real motion pattern in the real trajectory data set.But a moveing trajectory is usually predicted for the existing prediction method,which causes the prediction error to be large in this case.The second problem is that the trajectory points are all composed of two-dimensional latitude and longitude point,the discrepency between data is very small,which makes it hard to model the data directly using a single model.For the first problem,we proposes a mobile state prediction component to predict the state of moving objects.That is,the motion state is predicted first before the trajactory being predicted,hence the prediction accuracy can be effectively improved by this component.For the second problem,we use a multi-layer perceptron and a long short-term memory network to solve the problem.The multi-layer perceptron is used to capture the local features in the trajectory data,and the long shortterm memory network is used to capture the time dependence in the trajectory data.We combined the mobile state prediction component into a multi-layer perceptron and a lon short-term memory network to obtain a new random forest and MLP-LSTM based model of trajectory prediction.We verifies our proposed hybrid model in three real data sets.The experimental results show that the proposed hybrid model is superior to the existing comparison method on the three datasets.