Localization Under Consistent Assumptions Over Dynamics

TL;DR

The paper tackles localization under realistic dynamics by distinguishing movability from motion and introducing three dynamic classes: Static $S$, Semi-static $E$, and Dynamic $D$. It combines semantic segmentation with background-subtraction based filters to partition measurements and construct maps and localization pipelines that preserve the static-world assumption where appropriate. Empirical evaluation on the Oxford Radar RobotCar dataset demonstrates that using static measurements for mapping and/or localization, and applying filters to exclude semi-static/dynamic data, reduces localization error and variance, with the best results achieved when both map and localization are consistent with over dynamics. The approach implies that more accurate and robust robot localization in urban settings with semi-static objects can be achieved by explicitly modeling dynamic properties, with potential benefits for mapping and planning as well.

Abstract

Accurate maps are a prerequisite for virtually all mobile robot tasks. Most state-of-the-art maps assume a static world; therefore, dynamic objects are filtered out of the measurements. However, this division ignores movable but non-moving -- i.e., semi-static -- objects, which are usually recorded in the map and treated as static objects, violating the static world assumption and causing errors in the localization. This paper presents a method for consistently modeling moving and movable objects to match the map and measurements. This reduces the error resulting from inconsistent categorization and treatment of non-static measurements. A semantic segmentation network is used to categorize the measurements into static and semi-static classes, and a background subtraction filter is used to remove dynamic measurements. Finally, we show that consistent assumptions over dynamics improve localization accuracy when compared against a state-of-the-art baseline solution using real-world data from the Oxford Radar RobotCar data set.

Figures (4)

• Figure 1: Treating semi-static objects as static ones violates the static world assumption and causes mismatches between measurements and a map. In the figure, the map contains parked cars that have since moved away. When the observer returns, a new parked car is detected offset from the ones on the map. This causes matching errors, especially when the difference in poses is small or other features in the direction of the error are lacking or sparse.
• Figure 2: Overview of the proposed method. The mapping process is depicted in the first row, and the localization process is depicted in the second row. First, in i) and v), the dynamic classes for each measurement point are estimated. Then, the dynamic and semi-static objects, shown in red, are removed from the measurements used for building a map ii) and localizing vi). This results in measurements where only the reliable static measurements are used, shown in blue in iii) and vii). This allows the localization using consistent assumptions over dynamics between the map iv) and the measurement viii). This is in contrast to the state-of-the-art localization methods, where measurements are directly compared to the map, both containing semi-static and dynamic objects, causing localization errors.
• Figure 3: The experiment results. In the figure, the sample median is presented with a red line, and the blue box represents the range between 25$^{\text{th}}$ and 75$^{\text{th}}$ percentile, i.e., the interquartile range. The black dashed line presents the interval between the minimum and the maximum samples. Values over 1.5 times the interquartile range are marked as outliers, and displayed with a red plus symbol.
• Figure 4: A parking lot is a common scene where a large portion of the measurements originate from unreliable semi-static landmarks, shown in red in the baseline map. The configuration of the parking lot changes frequently, as seen from the maps created on different days (a)-(c). When the same measurements are filtered, only reliable static landmarks remain in the static map, as shown in (d).