[Paper Review] GRIP++: Enhanced Graph-based Interaction-aware Trajectory Prediction for Autonomous Driving

Xin Li, Xiaowen Ying+1 more|arXiv (Cornell University)|Jul 17, 2019

Autonomous Vehicle Technology and Safety40 references87 citations

TL;DR

GRIP++ extends GRIP by combining fixed and trainable graphs to capture inter-agent interactions for better trajectory prediction in urban driving, achieving state-of-the-art results on ApolloScape and faster runtime than CS-LSTM.

ABSTRACT

Despite the advancement in the technology of autonomous driving cars, the safety of a self-driving car is still a challenging problem that has not been well studied. Motion prediction is one of the core functions of an autonomous driving car. Previously, we propose a novel scheme called GRIP which is designed to predict trajectories for traffic agents around an autonomous car efficiently. GRIP uses a graph to represent the interactions of close objects, applies several graph convolutional blocks to extract features, and subsequently uses an encoder-decoder long short-term memory (LSTM) model to make predictions. Even though our experimental results show that GRIP improves the prediction accuracy of the state-of-the-art solution by 30%, GRIP still has some limitations. GRIP uses a fixed graph to describe the relationships between different traffic agents and hence may suffer some performance degradations when it is being used in urban traffic scenarios. Hence, in this paper, we describe an improved scheme called GRIP++ where we use both fixed and dynamic graphs for trajectory predictions of different types of traffic agents. Such an improvement can help autonomous driving cars avoid many traffic accidents. Our evaluations using a recently released urban traffic dataset, namely ApolloScape showed that GRIP++ achieves better prediction accuracy than state-of-the-art schemes. GRIP++ ranked #1 on the leaderboard of the ApolloScape trajectory competition in October 2019. In addition, GRIP++ runs 21.7 times faster than a state-of-the-art scheme, CS-LSTM.

Motivation & Objective

• Motivate improved trajectory prediction for autonomous driving in urban environments where interactions are complex.
• Propose GRIP++ to incorporate both fixed and trainable graphs to better model inter-agent interactions.
• Design a graph convolutional model followed by a Seq2Seq GRU-based trajectory predictor with residual connections.
• Evaluate GRIP++ on highway (NGSIM) and urban (ApolloScape) datasets to demonstrate accuracy gains and speedups.

Proposed method

• Represent scenes as graphs with fixed and trainable adjacency to model inter-agent interactions.
• Apply 2D 1x1 convolution to expand feature channels from input velocities.
• Use alternating graph operation and temporal convolution layers to learn spatial-temporal features.
• Employ a multi-block Seq2Seq trajectory predictor with encoder-decoder GRUs and residual connections to predict future positions.
• Train end-to-end with Adam; predict velocity deltas and convert to absolute positions.

Experimental results

Research questions

• RQ1 How does combining fixed and trainable graphs affect trajectory prediction accuracy in urban scenarios?
• RQ2 Can GRIP++ outperform state-of-the-art methods on urban trajectory datasets like ApolloScape?
• RQ3 Does utilizing velocity inputs and residual connections improve long-horizon predictions?
• RQ4 What is the impact of a neighborhood distance threshold on prediction performance?
• RQ5 How does GRIP++ compare in speed to CS-LSTM while maintaining accuracy?

Key findings

• GRIP++ achieves higher accuracy than state-of-the-art methods on ApolloScape trajectory data, surpassing TrafficPredict and StarNet while improving ADE/FDE metrics.
• On NGSIM datasets, GRIP++ provides competitive results and shows strong short-term and long-term prediction performance.
• GRIP++ runs 21.7x faster than CS-LSTM in urban trajectory prediction experiments.
• In ablations, using a fixed neighborhood threshold (D_close = 25 feet) and including nearby objects improves prediction accuracy, while expanding the neighborhood beyond this can degrade performance.
• The model predicts trajectories for all observed objects simultaneously, offering efficiency advantages over object-by-object prediction schemes.