Data Filtering Method for Intelligent Vehicle Shared Autonomy Based on a Dynamic Time Warping Algorithm

Zhenhai Gao, Tong Yu, Tianjun Sun, Haoyuan Zhao2022 · DOI

TL;DR

This paper tackles the massive data storage and processing burden in autonomous driving development by proposing a novel data filtering method. By using Dynamic Time Warping (DTW) to identify significant discrepancies between human driver and autonomous system control commands, the method intelligently records only the most informative data, drastically reducing storage needs and improving efficiency.

What Problem Does This Paper Address?

The rapid advancement of intelligent vehicles and shared autonomy has led to an explosion of driving data. While this "big data" is crucial for training and validating complex autonomous driving algorithms, it presents a significant challenge: the sheer volume of data makes efficient extraction, recording, and processing incredibly difficult and costly. Traditional methods often record vast amounts of redundant or low-quality data, slowing down algorithm development and increasing storage and labeling expenses. A key problem is identifying and capturing the critical data – specifically, scenarios where the human driver's intentions and actions significantly differ from those of the autonomous driving system. These discrepancies are vital for understanding and improving the system's performance in challenging or unexpected situations, but they are hard to pinpoint within the massive datasets.

Methods

The core of this paper is a Discrepancy Trigger Model that leverages the Dynamic Time Warping (DTW) algorithm. Here's how it works:

  • Focus on Control Commands: Instead of analyzing raw sensor data, the method focuses on the control commands (e.g., acceleration/deceleration, steering wheel angle) issued by both the human driver and the autonomous driving system.
  • Dynamic Time Warping (DTW): DTW is used to compare sequences of control commands over a specific time window. Unlike simpler methods like Euclidean distance, DTW is adept at handling time series data that are not perfectly synchronized, which is common with driver and system control cycles. It finds the optimal alignment between two sequences to measure their similarity or dissimilarity.
  • Discrepancy Quantification: The DTW algorithm calculates a "warping cost" that quantifies the difference between the driver's and the autonomous system's control command sequences.
  • Trigger Threshold: A statistically derived threshold is set for this warping cost. If the calculated warping cost exceeds this threshold, it signifies a significant discrepancy, triggering the recording of the relevant data segment.
  • Time Window Recording: A sliding time window approach ensures that when a discrepancy is detected, not just the instantaneous command but a preceding segment of data (e.g., 30 seconds before the trigger) is recorded, providing context.

Datasets

The paper describes experiments conducted using:

  • Simulation Data: Data was generated from simulations involving a passenger car equipped with ROS Kinetic and MATLAB.
  • Real-World Vehicle Data: Tests were performed on a real vehicle platform (Cincoze DS1201) equipped with ROS Kinetic, with the discrepancy trigger model operating in real-time on the onboard IPC.
  • Specific Algorithms Tested: The method was tested against an L2-level Adaptive Cruise Control (ACC) algorithm and an L4-level autonomous driving algorithm on a closed campus.

The paper does not introduce a new benchmark dataset but rather evaluates its method on data generated from these simulated and real-world driving scenarios.

Key Insights and Ideas

  • Control Command Discrepancy as a Key Indicator: The paper highlights that the most valuable data for improving autonomous driving systems isn't just any driving data, but specifically data where the system's intended actions diverge significantly from human driver behavior. This is where the system is likely to fail or perform suboptimally.
  • DTW's Suitability for Asynchronous Control Data: The choice of DTW is crucial. It elegantly handles the inherent time-axis misalignment between human driver inputs and the system's control outputs, which is a common challenge in this domain.
  • Intelligent Data Filtering Saves Resources: By focusing on discrepancies rather than just recording everything, the method offers a practical solution to the data storage and processing bottleneck, making development cycles more efficient.
  • Independent Longitudinal and Lateral Control Analysis: The approach treats longitudinal (acceleration/speed) and lateral (steering) control commands separately, acknowledging their distinct characteristics and control mechanisms.

Key Results

The proposed DTW-based discrepancy trigger model demonstrated significant improvements:

  • Data Reduction: On average, the model reduced recorded invalid data by 75.35%.
  • Storage Savings: This translates to saving approximately 2.65 TB of data storage space per hour for an autonomous driving system generating 3600 GB per hour.
  • Comparison to Instantaneous Triggering: Compared to simpler instantaneous value triggering methods, the DTW model saved an average of 307 GB of storage space per hour, indicating a more precise and effective filtering mechanism.
  • Qualitative Findings: The model was shown to accurately detect and record situations where control commands had distinct or even opposite trends, particularly in the lateral direction, and its triggering was more precise than instantaneous methods.

Limitations and Future Work

  • Independent Longitudinal and Lateral Analysis: The paper acknowledges that longitudinal and lateral control commands are often coupled in real-world driving. Computing discrepancies independently might lead to inaccuracies.
  • Future Work: The authors plan to explore a fusion of lateral and longitudinal discrepancy detection to enhance the model's accuracy in identifying differences in driving intentions. This integration aims to further improve the quality and effectiveness of data collection.

Summary updated

BibTeX
@misc{gao2022data,
  author = {Zhenhai Gao and Tong Yu and Tianjun Sun and Haoyuan Zhao},
  title = {{Data Filtering Method for Intelligent Vehicle Shared Autonomy Based on a Dynamic Time Warping Algorithm}},
  year = {2022},
  doi = {10.3390/s22239436},
}
Figure