Exploring Dynamic Context for Multi-path Trajectory Prediction

Abstract

To accurately predict future positions of different agents in traffic scenarios is crucial for safely deploying intelligent autonomous systems in the real-world environment. However, it remains a challenge due to the behavior of a target agent being affected by other agents dynamically, and there being more than one socially possible paths the agent could take. In this paper, we propose a novel framework, named Dynamic Context Encoder Network (DCENet). In our framework, first, the spatial context between agents is explored by using selfattention architectures. Then, two LSTM encoders are trained to learn temporal context between steps by taking the observed trajectories and the extracted dynamic spatial context as input, respectively. The spatial-temporal context is encoded into a latent space using a Conditional Variational Auto-Encoder (CVAE) module. Finally, a set of future trajectories for each agent is predicted conditioned on the learned spatial-temporal context by sampling from the latent space, repeatedly. DCENet is evaluated on the largest and most challenging trajectory forecasting benchmark Trajnet and reports a new state-of-theart performance. It also demonstrates superior performance evaluated on the benchmark InD for mixed traffic at intersections. A series of ablation studies are conducted to validate the effectiveness of each proposed module. Our code is available at https://github.com/wtliao/DCENet.

The Stanford Drone Dataset is More Complex than We Think: An Analysis of Key Characteristics

“…Vehicles make up a significant minority, and many of them are parked. In comparison, the inD dataset contains a larger percentage of moving cars but fewer pedestrians and classes.This suggests that the SDD may be more suited for applications that focus on pedestrians [3,31] while the inD dataset is suited for models intended for use in environments with both pedestrians and cars [32] This idea is further supported by differences in scene diversity and navigational behavior.…”

SCOUT: Socially-COnsistent and UndersTandable Graph Attention Network for Trajectory Prediction of Vehicles and VRUs

“…However, for inD benchmark we found better results when predicting directly x,y-coordinates. In Table II SCOUT is compared to other baselines and existing solutions using InD dataset following the same data preprocessing strategy as [19] in order to make a fair comparison. T obs = 3.2s and T pred = 4.8s are used, being the time interval 0.4s, i.e.…”

“…• Attentive Maps Encoder Network (AMENet) [25]: a generative model based on a conditional variational auto-encoder (CVAE) that uses attentive dynamic maps for interaction modeling. • Dynamic Context Encoder Network (DCENet) [19]:…”

Predicting Vehicles Trajectories in Urban Scenarios with Transformer Networks and Augmented Information

“…Recently, the STAR model [25] proposed to combine GATs to model spatial interactions, with Transformers to model temporal interactions. Finally, we can find recent proposals based on the combination of some encoder-decoder architecture with Conditional Variational Auto-Encoders (CVAE) such as AMENet [2] or DCENet [3]. CVAEs are used to encode spatial-temporal information into a latent space.…”

“…It is reasonable to think that agents are affected in their behavior by traffic conditions and road structure, so any potential solution must be flexible enough to be applicable to various scene contexts. In addition, although recent approaches model the behaviors of multiple agent types within a single model (vehicles, cyclists and pedestrians) [2], [3], having specific models for each agent type simplifies the problem, and facilitates the use of simple and effective architectures, such as Transformer (TF) networks. These have been proposed for Natural Language Processing (NLP) to deal with word sequences, using attention instead of sequential processing [4].…”