[Paper Review] Data Driven Prediction Architecture for Autonomous Driving and its Application on Apollo Platform

Kecheng Xu, Xiangquan Xiao+2 more|arXiv (Cornell University)|Jun 11, 2020

Autonomous Vehicle Technology and Safety17 references4 citations

TL;DR

This paper presents a fully automated, data-driven prediction pipeline for autonomous driving, deployed on Baidu's Apollo platform, that enables scalable, efficient model adaptation across diverse geographic regions. By integrating automated data annotation, feature extraction, hyperparameter tuning via Siamese networks, and support for both direct and indirect trajectory prediction, the system achieves up to 400% efficiency gain in parameter tuning compared to manual methods.

ABSTRACT

Autonomous Driving vehicles (ADV) are on road with large scales. For safe and efficient operations, ADVs must be able to predict the future states and iterative with road entities in complex, real-world driving scenarios. How to migrate a well-trained prediction model from one geo-fenced area to another is essential in scaling the ADV operation and is difficult most of the time since the terrains, traffic rules, entities distributions, driving/walking patterns would be largely different in different geo-fenced operation areas. In this paper, we introduce a highly automated learning-based prediction model pipeline, which has been deployed on Baidu Apollo self-driving platform, to support different prediction learning sub-modules' data annotation, feature extraction, model training/tuning and deployment. This pipeline is completely automatic without any human intervention and shows an up to 400\% efficiency increase in parameter tuning, when deployed at scale in different scenarios across nations.

Motivation & Objective

• To address the challenge of scaling autonomous driving prediction models across diverse geographic regions with varying traffic rules, terrain, and driving behaviors.
• To reduce manual intervention in model development and deployment by automating data annotation, feature extraction, training, and hyperparameter tuning.
• To enable efficient adaptation of prediction models for both direct trajectory prediction and intention-based post-trajectory generation across different scenarios.
• To improve deployment efficiency and reduce costs in large-scale autonomous vehicle operations across multiple nations and vehicle platforms.

Proposed method

• The pipeline consists of an onboard module (message pre-processing, model inference, trajectory post-processing) and an offboard data pipeline (automatic data annotation, feature extraction, model training, hyperparameter auto-tuning, result evaluation).
• For direct prediction, a semantic map and LSTM-based model are trained end-to-end with automated data processing, achieving performance comparable to manually-tuned models.
• For indirect prediction, an intention classification model (e.g., lane change, turn) is combined with a lattice-based post-trajectory generator that uses cost functions with learnable weights.
• A Siamese network is used to automatically learn optimal cost function weights (θ₁, θ₂, θ₃) in Equation (4) by minimizing the loss in Equation (5), which compares candidate trajectory costs to ground-truth trajectories.
• The objective function in Equation (5) uses a margin loss to ensure ground-truth trajectories have lower cost than candidate trajectories, with δ as a small positive margin to prevent degenerate solutions.
• The system supports both direct and indirect prediction workflows with minimal human intervention, enabling rapid deployment in new geo-fenced areas.

Experimental results

Research questions

• RQ1 How can a prediction pipeline be automated to reduce human intervention in data annotation, model training, and hyperparameter tuning for autonomous driving?
• RQ2 What is the impact of automated hyperparameter tuning via Siamese networks on deployment efficiency across different geographic regions?
• RQ3 Can a fully automated pipeline achieve comparable prediction performance to manually tuned models in both direct and indirect trajectory prediction?
• RQ4 How does the integration of learned cost weights in post-trajectory generation improve adaptability to regional driving patterns and traffic rules?
• RQ5 To what extent can this pipeline scale across diverse vehicle platforms and global deployment scenarios?

Key findings

• The proposed data-driven prediction pipeline achieved up to a 400% increase in efficiency for hyperparameter tuning compared to manual methods when deploying in new geo-fenced areas.
• The automated pipeline enabled deployment of prediction models across different nations and vehicle platforms with minimal human intervention, significantly reducing operational costs.
• The Siamese network-based auto-tuning method successfully learned region-specific cost function weights (θ₁, θ₂, θ₃) that improved trajectory selection accuracy without manual calibration.
• The direct prediction model using semantic maps and LSTM achieved performance comparable to manually trained models, despite full automation of data processing and training.
• The indirect prediction pipeline with intention classification and post-trajectory generation demonstrated robustness to regional variations in driving behavior and traffic rules.
• The system was successfully deployed on multiple fleets of self-driving vehicles in both China and the US, validating its real-world scalability and adaptability.