Automated synchronization of driving data using vibration and steering events☆
Introduction
Large multi-sensor on-road driving datasets offer the promise of helping researchers develop a better understanding of driver behavior in the real world and aid in the design of future advanced driver assistance systems (ADAS) [1], [2]. As an example, the Strategic Highway Research Program (SHRP 2) Naturalistic Driving Study includes over 3400 drivers and vehicles with over 5,400,000 trip records [3] that contains video, telemetry, accelerometer, and other sensor data. The most interesting insights are likely to be discovered not in the individual sensor streams but in their fusion. However, sensor fusion requires accurate sensor synchronization. The practical challenge of fusing “big data”, especially in the driving domain, is that it is often poorly synchronized, especially when individual sensor streams are collected on separate hardware [4]. A synchronization error of 1 s may be deemed acceptable for traditional statistical analyses that focus on data aggregated over a multi-second or multi-minute windows. But in the driving context, given high speed and close proximity to surrounding vehicles, a lot can happen in less than one second. We believe that the study of behavior in relation to situationally relevant cues and the design of an ADAS system that supports driver attention on a moment-to-moment basis requires a maximum synchronization error of 100 ms. For example, events associated with glances (e.g., eye saccades, blinks) often occur on a sub-100-ms timescale [5].
Hundreds of papers are written every year looking at the correlation between two or more aspects of driving (e.g., eye movement and steering behavior). The assumption in many of these analyses is that the underlying data streams are synchronized or aggregated over a long enough window that the synchronization error is not significantly impacting the interpretation of the data. Often, these assumptions are not thoroughly tested. The goal of our work is to motivate the feasibility and the importance of automated synchronization of multi-sensor driving datasets. This includes both “one-factor synchronization” where the passive method is the primary synchronizer and “two-factor synchronization” where the passive method is a validator of a real-time clock based method engineered into the data collection device.
For the passive synchronization process, we use two event types: (1) vehicle vibration and (2) vehicle steering. These event types can be detected by video, audio, telemetry, and accelerometer sensors. Cross-correlation of processed sensor streams is used to compute the time-delay of each sensor pair. We evaluate the automated synchronization framework on a small dataset and achieve an average synchronization error of 13 ms. We also characterize the increase in accuracy with respect to increasing data stream duration which motivates the applicability of this method to online synchronization.
The implementation tutorial and source code for this work is available at: http://lexfridman.com/carsync.
Section snippets
Related work
Sensor synchronization has been studied thoroughly in the domain of sensor networks where, generally, a large number of sensor nodes are densely deployed over a geographic region to observe specific events [6], [7]. The solution is in designing robust synchronization protocols to provide a common notion of time to all the nodes in the sensor network [8]. These protocols rely on the ability to propagate ground truth timing information in a master-slave or peer-to-peer framework. Our paper
Dataset and sensors
In order to validate the proposed synchronization approach we instrumented a 2014 Mercedes CLA with a single-board computer and consumer-level inexpensive sensors: 3 webcams, a shotgun microphone behind the rear tire, GPS, an IMU module, and a CAN controller for vehicle telemetry. The instrumented vehicle is shown in Fig. 1. Details on the positioning of the sensors are provided in the figure’s caption. Through empirical testing we found that small changes in the position of the sensors did not
Synchronization framework
Unless otherwise noted, the figures in this section show sensor traces and cross correlation functions for a single 37 min example run. The two synchronizing event types are vibrations and steering, both densely represented throughout a typical driving session.
Conclusion
Analysis and prediction based on fusion of multi-sensor driving data requires that the data is synchronized. We propose a method for automated synchronization of vehicle sensors based on vibration and steering events. This approach is applicable in both an offline context (i.e., for driver behavior analysis) and an online context (i.e., for real-time intelligent driver assistance). We show that a synchronization error of 13.5 ms can be achieved for a driving session of 35 min.
Acknowledgment
Support for this work was provided by the New England University Transportation Center, and the Toyota Class Action Settlement Safety Research and Education Program. The views and conclusions being expressed are those of the authors, and have not been sponsored, approved, or endorsed by Toyota or plaintiffs class counsel.
References (17)
- et al.
Semi-automated annotation of discrete states in large video datasets
AAAI
(2016) Driver assistance systems and the transition to automated vehicles: a path to increase older adult safety and mobility?
Public Policy Aging Rep.
(2014)Design of the in-vehicle driving behavior and crash risk study: in support of the SHRP 2 naturalistic driving study
(2011)- et al.
Reliability meets big data: Opportunities and challenges
Qual. Eng.
(2014) - et al.
Time on task and blink effects on saccade duration
Ergonomics
(1996) - et al.
Time synchronization in sensor networks: a survey
Netw., IEEE
(2004) - et al.
Clock synchronization in wireless sensor networks: an overview
Sensors
(2009) - et al.
Fine-grained network time synchronization using reference broadcasts
ACM SIGOPS Operat. Syst. Rev.
(2002)
Cited by (32)
Automated vehicle data pipeline for accident reconstruction: New insights from LiDAR, camera, and radar data
2023, Accident Analysis and PreventionComprehensive driver behaviour review: Taxonomy, issues and challenges, motivations and research direction towards achieving a smart transportation environment
2022, Engineering Applications of Artificial IntelligenceCitation Excerpt :Analysing the environment surrounding a driver can be used to improve ADAS reaction, hence an enhancement of passenger comfort ability (Klimke et al., 2014). Therefore, the current method of implementing ADAS should also be improved (Fridman et al., 2016). ADAS particularly helps to reduce and avoid collisions, improve road safety, reduce energy consumption and mitigate emissions (Wang et al., 2015a).
Data synchronization by continuous spatial measurement with Double Ballbar
2021, Measurement: Journal of the International Measurement ConfederationCitation Excerpt :In respect of the mechanical systems, some researches in recent years have purposed to exploit the vibration in signals to synchronize them. For instance, L. Fridmann et al. determined the time delay in each vehicle sensor by calculating the cross correlation function between data streams [32]. They found out that pairing all sensors with the accelerometer delivers the best results, since the accelerometer captures the vibration signal best.
E-Scooter safety: The riding risk analysis based on mobile sensing data
2021, Accident Analysis and PreventionIdentifying baseflow source areas using remotely sensed and ground-based hydrologic data
2024, Hydrological ProcessesRoad Condition Detection Techniques: A Comparative Analysis based on Supervised Machine Learning Techniques
2023, 14th International Conference on Advances in Computing, Control, and Telecommunication Technologies, ACT 2023
- ☆
This paper has been recommended for acceptance by R. Davies.