Short-Term 4D Trajectory Prediction for UAV Based on Spatio-Temporal Trajectory Clustering

Abstract

To improve the trajectory prediction accuracy of unmanned aerial vehicles (UAVs) with random behavior intentions, this paper presents a short-term four-dimensional (4D) trajectory prediction method based on spatio-temporal trajectory clustering. A spatio-temporal trajectory clustering algorithm is first designed to cluster the UAV trajectory segments divided by a fixed time window. Each trajectory segment is given a category label that represents some certain type of behavior characteristics, such as climbing, turning, descending, etc. The convolutional neural network (CNN) is used to identify the category label of a given trajectory segment by learning the behavior characteristics of different trajectory segments. Based on the long-short-term memory network (LSTM), a short-term trajectory prediction model for different categories of label trajectory segments is established. The global trajectory prediction includes several steps adopting the corresponding prediction models. Historical trajectory data of UAVs are used to validate the proposed prediction method. Experiment results indicate that the method can obtain obviously better prediction accuracy in a short prediction time range (0-3s) with acceptable efficiency compared to LSTM, GRU and velocity trend extrapolation.

Traffic State Estimation with A multi-head Attention-based Transformer by Spatio-Temporal Autoencoding Transformer (STAT Model)

“…Because of their special design in the gating mechanism, LSTM models, which are a subset of RNNs, can deal with the issue of vanishing gradients to control its memory in RNN networks during training and hence overcomes the long-term dependence of learning [29,30]. The GRU model, which is referred to as a sister of the LSTM model [7], was proposed by Chung et al [31] for the first time, and [32] was one of the pioneers of applying GRU in traffic state estimation on PeMS datasets demonstrating the superior convergence speed of GRU compared to LSTM. In any case, these RNN-based algorithms cannot extract spatial characteristics adaptively, and geographical data must be separately encoded into the input.…”

“…Gated recurrent unit (GRU), long short-term memory (LSTM), stacked autoencoders (SAEs) in complex non-linear problems and the automatic extraction of spatial-temporal features in large-scale data, have recently been explored to address some of these aforementioned issues [6]. Conventional deep learning methods in sequence modeling have difficulties in extracting long-term dependencies in data sequences [7,8]. In the realm of modeling long sequential data, researchers have commonly acknowledged the significance of CNN and recurrent neural network (RNN) in sequence-tosequence (seq2seq) modeling, especially when dealing with sequences of varying lengths.…”

Distributed Mobility Management Support for Low-Latency Data Delivery in Named Data Networking for UAVs

“…Subsequently, the prod packet, which includes the producer's predicted coordinates, to notify Several methods have been proposed to predict the future locatio prediction methods aim to estimate the UAV's position at a future tim rent location and other relevant factors. Common approaches include lation [34], sensor fusion [35], Kalman filtering [36], machine learning models [38]. In this paper, a Kalman Filter (KF) [39] method is used wi Positioning System (GPS) and Inertial Measurement Unit (IMU) sens predicting the UAV's future position coordinate.…”

“…These prediction methods aim to estimate the UAV's position at a future time based on its current location and other relevant factors. Common approaches include trajectory extrapolation [34], sensor fusion [35], Kalman filtering [36], machine learning [37], and dynamic models [38]. In this paper, a Kalman Filter (KF) [39] method is used with the aid of Global Positioning System (GPS) and Inertial Measurement Unit (IMU) sensor technologies for predicting the UAV's future position coordinate.…”

Hierarchical Clustering Algorithm for Multi-Camera Vehicle Trajectories Based on Spatio-Temporal Grouping under Intelligent Transportation and Smart City

“…The statistical model method uses the Gaussian mixture model [ 40 ] and other statistical models to construct the trajectory category probability function to obtain the trajectory sample category. In addition to the above methods, in recent years, with the continuous extension of deep learning, there are a large number of deep learning algorithms [ 41 , 42 ]. Semi-supervised trajectory clustering algorithm [ 43 , 44 ] fully combines supervised clustering algorithm with unsupervised clustering algorithm to reduce the label labor cost and minimize the over-fitting problem in the clustering process.…”