Overview
For automakers working to bring autonomous driving to public roads, safe operation depends heavily on sensor systems. LiDAR and long-range 3D vision sensors are two common ranging solutions, with notable differences in performance under adverse weather and road conditions.
Nodar, a supplier of stereo vision technology for autonomous vehicles, recently ran head-to-head performance tests to compare how LiDAR and stereo cameras handle low light, darkness, and adverse weather, and how well they detect small obstacles on the road. In each test, vehicles equipped with a high-performance LiDAR system were compared to a wide-baseline stereo sensor using 5.4-megapixel cameras and 30° field-of-view lenses.
The results showed that in adverse weather and low-light conditions, 3D stereo vision significantly outperformed LiDAR, conditions that are critical for autonomous vehicle safety. Coupled with stereo sensors' known daytime performance, these results address concerns about camera-based sensor performance and support stereo vision as a primary 3D sensor for L3+ driving applications.
Tests in Rain and Fog
In November 2022, tests were performed in a vehicle environmental chamber in Germany that simulated rain and fog of varying intensity under daytime and nighttime lighting. On dry roads, both LiDAR and stereo cameras returned 100% valid range data per measurement point. However, the point-cloud density from the high-resolution stereo camera was 50x higher than the LiDAR point cloud.
Detailed results by condition:
- Heavy rain (32 mm/h): Stereo vision performance dropped slightly, accurately measuring about 95% of the scene. LiDAR performance fell noticeably, measuring less than 80% of the scene.
- Torrential rain (96 mm/h): The camera-based sensor recorded accurate readings on approximately 70% of data points. LiDAR's ability to return valid range data fell below 40%, meaning over 60% of distance measurements were lost or invalid.
- Fog (45 m visibility): Stereo vision returned nearly 70% accurate distance measurements and could detect objects that were difficult or impossible to see by eye in the fog. LiDAR performance declined to about 20% accurate distance measurements, meaning 80% of distances were lost or invalid.
Night Driving
Night driving tests were completed in October 2022 on roads around Boston. These tests compared the number of valid range measurements returned by the wide-baseline stereo camera and LiDAR at illumination levels from full daylight (10,000 lux) down to night (1 lux).
LiDAR returned roughly 600,000 data points per second across all illumination levels. In contrast, the stereo sensor's returned point count varied with light level as follows:
- Sunrise, sunset, full daylight, and overcast (1,000–10,000 lux): 40 million points per second
- Overcast sunrise and sunset (100–1,000 lux): over 30 million points per second
- Urban light pollution (10–100 lux): about 20 million points per second
- Moonlight (1 lux): over 10 million points per second
Nighttime Obstacle Detection
In April 2023, nighttime obstacle detection tests were performed at a closed small airport in Maine, where dark skies minimized light pollution. A wide-baseline stereo system was installed on a passenger vehicle and tested using high beams and low beams.
Night detection results for wood, human mannequins, and traffic cones:
- Wood (12 cm high): Stereo detected a piece of wood on the road at 130 m with high beams and 100 m with low beams. LiDAR detected the wood out to 50 m.
- Adult-size mannequin, lying down (30 cm high): Stereo detected a lying mannequin at 160 m with high beams and 100 m with low beams. LiDAR detected the mannequin at 100 m.
- Child-size mannequin, standing (100 cm high): Stereo detected a standing mannequin at 200 m with high beams and 100 m with low beams. LiDAR detected the mannequin at 100 m.
- Traffic cone (70 cm high): Stereo detected a traffic cone at 200 m with high beams and 160 m with low beams. LiDAR detected the cone out to 50 m.
Conclusion
The test series indicates that 3D stereo vision outperforms LiDAR across a range of adverse conditions. Although LiDAR can produce accurate distance measurements, the tested multi-camera stereo system produced point-cloud densities several orders of magnitude higher than the LiDAR, which reduces LiDAR's ability to detect small objects. The tests also show that LiDAR performance degrades more severely than stereo vision in low-visibility conditions. Finally, despite common assumptions that camera-based systems perform poorly at night, the wide-baseline stereo sensor in these tests measured object distances to ranges approximately twice those of the LiDAR sensor.
