Fast Trajectory Prediction Method With Attention Enhanced SRU

Abstract

LSTM (Long-short Term Memory) is an effective method for trajectory prediction. However, it needs to rely on the state value of the previous unit when calculating the state value of neurons in the hidden layer, which results in too long training time and prediction time. To solve this problem, we propose Fast Trajectory Prediction method with Attention enhanced SRU (FTP-AS). Firstly, we devise an SRU (Simple Recurrent Units) based trajectory prediction method. It removes the dependencies on the hidden layer state at the previous moment, and enables the model to perform better parallel calculation, speeding up model training and prediction. However, each unit of the SRU calculates the state value at each moment independently, ignoring the timing relationship between the track points and leading to accuracy decrease. Secondly, we develop the attention mechanism to enhance SRU. The influence weight for selective learning is gained by calculating the matching degree of the hidden layer state value at each moment to improve the accuracy of the prediction. Finally, experimental results on MTA bus data set and Porto taxi data set showed that FTP-AS was 3.4 times faster and about 1.7% more accurate than the traditional LSTM method.

A Novel Lane-Change Decision-Making With Long-Time Trajectory Prediction for Autonomous Vehicle

“…LSTM plays an important role in trajectory prediction [34][35][36], however, its long training and prediction times have traditionally been a limitation. To address this issue, [37] proposed a novel attention-enhanced SRU from a cellular unit perspective, which has shown good performance in trajectory prediction. [38] has demonstrated that GRU is more efficient in training and achieves performance similar to LSTM in time-series data regression prediction.…”

Meta-Learning over Time for Destination Prediction Tasks

“…In the context of Porto, early successes were achieved by relatively simple multi-layer perceptron (MLP) models, such as the competition-winning result of De Brébisson et al [7], which outperformed competitors such as ensembles of regression trees [28]. More recent papers have predominantly employed recurrent networks instead [9,33,34,47,54,64], particularly long short-term memory (LSTM) networks [19]. Notable exceptions include Lv et al [36], who plot trajectories graphically and then model them as 2D images rather than GPS points or embeddings, using convolutional networks, and Tsiligkaridis et al [55], who use Transformers [57].…”

“…Competitors were challenged to predict a taxi's final GPS destination given only a limited "prefix" of the first 𝑛 GPS points and limited metadata about the taxi ride (notably not including the total length). More than 380 teams participated in the competition, drawing research interest that continued even after its conclusion [9,33,34,36,47,54,55,64,65]. The Porto dataset has also contributed to other geospatial modeling interests, including passenger demand modeling [31,41,46,48], travel time prediction [1,6,13,16,18,29], and unsupervised geospatial learning and anomaly detection [20,24,27,35,53].…”