Self-Assessment and Correction of Sensor Synchronization

TL;DR

The results show that the approach accurately estimates time offsets and, thus, is able to detect and assess synchronization issues and mitigates the effect of synchronization issues on subsequent modules in autonomous software stacks, such as tracking systems that heavily rely on accurate measurement time stamps.

Abstract

We propose an approach to assess the synchronization of rigidly mounted sensors based on their rotational motion. Using function similarity measures combined with a sliding window approach, our approach is capable of estimating time-varying time offsets. Further, the estimated offset allows the correction of erroneously assigned time stamps on measurements. This mitigates the effect of synchronization issues on subsequent modules in autonomous software stacks, such as tracking systems that heavily rely on accurate measurement time stamps. Additionally, a self-assessment based on an uncertainty measure is derived, and correction strategies are described. Our approach is evaluated with Monte Carlo experiments containing different error patterns. The results show that our approach accurately estimates time offsets and, thus, is able to detect and assess synchronization issues. To further embrace the importance of our approach for autonomous systems, we investigate the effect of synchronization inconsistencies in tracking systems in more detail and demonstrate the beneficial effect of our proposed offset correction.

Figures (6)

• Figure 1: The overall structure of our proposed self-assessment and correction approach is illustrated. It can mainly be divided into two parts, the offset estimation and the self-assessment and correction part. First, the offset estimation uses motion estimates of two sensors. A possible timestamp offset and a respective estimation uncertainty are estimated based on the rotational movements. Second, the self-assessment investigates the estimated offset over time. When deviations are detected, the self-assessment additionally decides, based on the uncertainty, whether timestamps can be corrected or data needs to be discarded. Respectively, sensor data is updated or, if necessary, discarded in the timestamp correction block before other modules further process it.
• Figure 2: The general motion of two sensors over time is shown. Sensor 1 (blue) and sensor 2 (red) are rigidly connected with the transformation $T$. $A_i$ and $B_i$ represent different positions of sensor 1 and sensor 2 at time step $i$, respectively; and $V^A_i$ and $V^B_i$ represent the motion of a sensor between two consecutive time steps. For each motion, the transformation cycle is marked with a gray circle.
• Figure 3: The effect of a measurement time offset on the tracking system is illustrated. A moving object (circle) is shown at different times. At time $t$, the object gets measured by a sensor (cross), but due to synchronization issues, the sensor incorrectly assigns the time step $t+dt$ to the measurement. Now, the tracking system would use this measurement to update the estimation of time $t+dt$. This introduces additional errors and violates the assumptions of the Kalman filter.
• Figure 4: The simulated path of the ego vehicle (red) and the target vehicle (blue). Both move on a plane, resulting in 2D motion.
• Figure 5: Results of different experiment settings are illustrated. For both ramp and steps error profiles (left/right) and two noise levels (top/bottom) the results are illustrated in the respective figure. In all figures, the ground truth time shift (red) and the estimated time shift (green) are plotted in the upper diagram. As a reference, the sensor rotation of the first simulation run is also shown in the background. For the estimated time shift, the green line represents the median value. The 25th and 75th quantiles are illustrated in gray. In the lower diagram, the absolute estimation error (red) and the estimation uncertainty (blue) are depicted.