Motion Prediction of Multi-agent systems with Multi-view clustering

TL;DR

This paper identifies clusters that incorporate the minimum cost-to-go function of a relevant optimal control problem as a metric for clustering between the groups among agents, where groups with similar associated costs are assumed to be likely to move together.

Abstract

This paper presents a method for future motion prediction of multi-agent systems by including group formation information and future intent. Formation of groups depends on a physics-based clustering method that follows the agglomerative hierarchical clustering algorithm. We identify clusters that incorporate the minimum cost-to-go function of a relevant optimal control problem as a metric for clustering between the groups among agents, where groups with similar associated costs are assumed to be likely to move together. The cost metric accounts for proximity to other agents as well as the intended goal of each agent. An unscented Kalman filter based approach is used to update the established clusters as well as add new clusters when new information is obtained. Our approach is verified through non-trivial numerical simulations implementing the proposed algorithm on different datasets pertaining to a variety of scenarios and agents.

Figures (9)

• Figure 1: Multi-agent scenario at a crosswalk with both pedestrian and vehicular agents. An ego agent would have to predict the behaviour of various types of other agents in the environment for safe navigation. Agents highlighted with the same color indicate potential group formations based on their similar motion pattern, proximity and intent.
• Figure 2: Example of complete linkage cluster-cluster grouping between two clusters $C_{I}$ and $C_{J}$. The shaded circles represent individual agents and the boundary indicates that they belong to one cluster. Complete linkage implies that the farthest agents of a group are similar. The clustering algorithm is applied to the agents connected by the solid black line in (c) to establish whether a complete linkage exists between the two clusters. The final cluster group is observed in (d).
• Figure 3: Prediction results on a scene from the Trajnet++ dataset. The circular markers represent the agent position predictions while the $'\times'$ markers represent the truth data for the position of the agent during the observed timeframe. The region enclosed by a red circle is the $3\sigma$ confidence ellipsoid around the cluster center.
• Figure 4: Prediction results on a pedestrian crossing scene from the Trajnet++ dataset. The circular markers represent the agent predictions while the $'\times'$ represent the truth data for the position of the agent during the observed time frame. The region enclosed by the red circle is the $3\sigma$ confidence ellipsoid around the cluster center.
• Figure 5: FDE and ADE calculated over several scenes for all the agents present on the scene. The average distance error in 'blue' is generally a lower value than the FDE in 'red' over the different scenarios.