In the domain of harsh-environment automation, particularly in autonomous mining and construction, robust and reliable sensing is critical. Conventional sensors, including Light Detection and Ranging (LiDAR) and vision-based systems (e.g., RGB cameras), primarily operate within the visible or near-visible light spectrum (400 nm to 2000 nm). These wavelengths, however, exhibit limited penetration through contaminants like dust, mud, and snow - a frequent problem in outdoor settings like mines and large scale construction sites. Such contamination significantly compromises sensor effectiveness within hours, posing a challenge to both reliability and safety in scenarios where operational downtime is infeasible.

Novel 4D imaging Radars however, operate at lower frequencies, e.g. 77GHz, demonstrating superior penetration through contaminant layers due to their millimeter-wave technology. Despite their seemingly erratic, noisy, and unpredictable scan patterns compared to LiDAR they are very robust and high-quality when coupled with specialised perception algorithms [2].

Figure 2: Comparative Analysis of LiDAR vs. Radar Scan Patterns. LiDAR provides geometric sampling of LiDAR the measurements vs. the localized reflectivity maxima of a Radar sensor [1].

At sensmore, we advocate for the superiority of 4D imaging Radars in heavy-duty automation, particularly given the anticipated degradation of vision-based systems in continuous operations.

To validate this premise, we present a comparative degradation analysis between automotive-grade LiDAR and a next-generation 77GHz imaging Radar sensor at one of our active mine installations.

Experiment Setup

Our study utilizes Komatsu785 [3] haulage trucks, fitted with Velodyne VLP32c LiDAR and 77GHz imaging Radar sensors directly next to each other, operating in a Western European outdoor mine. The sensors, positioned at 2m height, underwent routine mining operations for approximately 18 hours daily, including haulage, loading, and unloading.

Figure 3: One of the haulage trucks equipped with our system used for this analysis.

Conditions

The weather conditions were relatively stable with damp and cloudy, partially sunny weather. The wet conditions meant that the ground was usually muddy, with minimal dust in the air. The precipitation for the experiment days is given in Table 1.

Table 1: Weather data from a weather station roughly 10 km from the mine.

Methods

Sensor degradation was quantified through raw data analysis from both sensor types. Key metrics included measurement counts, range, and angular point distributions. Two distinct representations of sensor performance were created:

1. Range Distribution Plots: Daily counts of points within specific range intervals.
2. Azimuth-Range Histogram: Sorting of measurements into predefined bins, extracting the maximum range detected per day, focusing on the top 99% percentile to mitigate outliers.

Results

LiDAR

The LiDAR range distribution plot (Figure 4) shows very strong sensor degredation for each day of operation. Initial days showed measurements across all ranges up to 100m, while later days exhibited a strong reduction, particularly beyond 35m. Only a fraction of the measurements remain.

Figure 4: LiDAR Range Plot with Daily Widths Proportional to Mean Point Count per Frame. Almost all relevant Points are missing on Day 8.

Analyzing the 99th percentile range-azimuth histogram on day 1 vs. day 7 (Figure 5), it can be seen the physical sensing capability of the LiDAR system has dropped to 40m in the best case.

Figure 5: 99th percentile robust max. histogram of the LiDAR range over the azimuth bins.

Radar

In constrast the same analysis shows a very stable measurement capability for the Radar system (Figure 6) despite heavy contamination of the sensing surface with layers of mud.

Figure 6: Radar Range Plot with Daily Widths Proportional to Mean Point Count per Frame. The majority of points remain as on Day 1.

The same result can be seen by investigating the robustly measured maximum (99th percentile) range across the different azimuth bins as seen in Figure 7.

Figure 7: 99th percentile robust max. histogram of the Radar range over the azimuth bins.

Conclusion

The 4D imaging Radar maintained consistent mid and long-range measurement capabilities throughout the analysis, with no major degradation observed in higher ranges during the last two days. The overall measurement count remained stable, underscoring minimal impact on sensor performance.

Outlook

As evidenced by our analysis, it is evident that advanced imaging Radar sensors are poised to play a pivotal role in harsh environments like mining and construction. At sensmore, we are at the forefront of this technological shift, developing advanced automation solutions that leverage the strengths of this novel imaging Radar technology. Our innovative Radar foundational AI is specifically designed to address the unique challenges associated with this type of sensor, including issues related to noise, data sparsity, and the inherent complexity of the raw data. By leveraging the robust capabilities of imaging Radar, we aim to revolutionize the standard for environmental adaptability and reliability in heavy machinery automation.

References

[1] Igal Bilik. “Comparative Analysis of Radar and Lidar Technologies for Automotive Applications”. In: IEEE Intelligent Transportation System Magazine (2022).

[2] Zhou, Y.; Liu, L.; Zhao, H.; López-Benítez, M.; Yu, L.; Yue, Y. “Towards Deep Radar Perception for Autonomous Driving: Datasets, Methods, and Challenges”. Sensors 2022, 22, 4208. https://doi.org/10.3390/s22114208

[3] https://www.komatsu.eu/en/product-archive/rigid-dump-trucks/hd785-7