SA-LSTM: A Trajectory Prediction Model for Complex off-road Multi-agent Systems Considering Situation Awareness Based on Risk Field

Abstract

Autonomous Vehicles have wide-ranging applications in off-road environments. Off-road vehicular scenes can be abstracted as multi-agent systems, and trajectory prediction is a critical process for context understanding. Agents compete and cooperate other to pursue their individual targets, which makes trajectories complicated and changeable. Hence, in order to derive precise predictions, it is necessary to reason how agents interact with each other. However, current prediction algorithms lack a unified and appropriate method for spatiotemporal reasoning. Previous methods mainly rely on vehicle-lane constraints to capture features, which is only applicable in structured environments. To address this issue, this paper proposes a novel off-road multi-agent trajectory prediction framework called SA-LSTM. This framework comprises situation awareness extraction and Long Short Term Memory (LSTM) prediction backbones. Based on the analysis of agents' movement patterns, we decompose actions into the maintenance of previous movements and the strategy variations in response to environment situations. In situation awareness extraction, risk field and pooling layers are applied to filter interpretable awareness. As for prediction backbones, proper LSTM networks are selected to adapt the task features. In short-term predictions, a residual mechanism is used to preserve physical inertia, while for long-term predictions autoregression process is applied. We also establish a multi-agent dataset based on human-manipulated chase-and-run tasks to train and validate the performance of SA-LSTM. Experiments show that compared with the best baseline model at each prediction length, SA-LSTM reduces the mean absolute error by 7.27%, 32.99%, and 28.07% at the prediction time of 0.6, 3.0, and 6.0 seconds, respectively, proving better prediction accuracy.