Keywords

1 Automated Driving and LiDAR Sensor

Autonomous cars are at the forefront of developed automotive technologies. The driver assistance system called ADAS (Advanced Driver Assistance System) is being developed, and its elements are slowly becoming standard equipment for vehicles. Elements of this system, such as active cruise control, lane-keeping assistant, and automatic recognition of road signs, are commonly found in new vehicles. The purpose of introducing these technologies is primarily to increase the level of safety and relieve the driver [2]. Therefore, developing ADAS systems and sensors necessary for their proper operation has become a priority for both car manufacturers and scientists.

The more sophisticated systems in the autonomous vehicle use a range of more advanced sensors such as radar, cameras, and laser scanners (LiDAR). The use of LiDARs (Light Detection and Ranging) in autonomous cars have become popular because of reducing their cost, size and complexity. They provide reliable and detailed data and have a relatively long effective operating range (even up to 240m).

The general principle of a pulsed LiDAR sensor is to measure the time from the emission to the return of the laser beam reflected from the target and the energy of this beam. This process leads to the generation of data in the form of a point cloud, which contains information about the location and intensity of the given points (which results from the beam energy) [5]. At a later stage, based on this information, the perception system (using vision algorithms) of the autonomous car locates and creates an image of the surroundings, as well as detects and tracks moving objects. Although LiDAR sensors are considered to be highly reliable and precise, and the data they provide are less disturbed by bad weather conditions than, for example, in the case of a camera, so adverse weather conditions can also significantly reduce the effectiveness of their operation [3, 6, 7].

Due to the widespread use of LiDAR sensors and the fact that it is a very dynamically developing technology, it is necessary to test this type of device and check the capabilities of algorithms and artificial neural networks dedicated to object detection.

Two methods are used to test and validate LiDAR sensor performance:

  • simulation,

  • tests on test rigs (e.g. in special chambers) [11].

Simulation methods allow large amounts of data to be tested quickly. Suitable software is able to simulate rain or fog by e.g. adding artificial noise to the image. However, without a real test rig on which at least preliminary measurements can be made, validation of such simulations is very difficult because the calculated theoretical range of the LiDAR sensor depends on the coefficients, the values of which should be selected empirically and each time for a specific device. [8, 12]. Computer simulation of LiDAR operation during rain is even more difficult as objects of different intensities generate different drops in detection efficiency. Simulations, on the other hand, can work well in testing various detection algorithms based on data obtained through real experiments at test rigs.

2 The LiDAR Sensor Testing Methods

The test rig must allow for the most credible simulation of road conditions while ensuring the repeatability of these conditions. Another advantage is the ability to add new elements to the examined scene, which makes it possible to extend the scope of research. The construction of the position itself is not an easy task because it requires appropriate knowledge, space, time and financial resources. Many researchers decide to conduct research in a ready, specially prepared chamber, such as the CERMA Climate Chamber in Clermont, France [1, 11]. This chamber allows you to create a stable fog with specific parameters of air transparency (from 5 to 200 m). It also allows you to simulate rain taking into account the size of the droplets (two sizes) and different rainfall intensities. On Fig. 1 is shown the chamber from the Clermont Research Center.

Fig. 1.
figure 1

CERNA chamber. Source: [1].

Full size image

While the Clermont chamber’s ability to simulate atmospheric conditions is truly impressive, the chamber is not without some drawbacks. Simulating rain allows producing rain with an intensity of 10 mm/h to 150 mm/h. The classification of precipitation intensity assumes that precipitation above 7.5 mm/h should be considered strong [4]. Thus, even the exceptionally refined chamber in CERMA does not allow simulating light to medium rainfall.

Table 1. Characteristics of rain intensity. Source: [11].
Full size table

Table 1 shows the basic characteristics of rain intensity. Most rain simulators can only simulate strong rainfall.

The Virginia Smart Road could serve as an example of another complex testing rig. It is a 3.5 km long test section of the road that allows you to test many issues in the fields of mechanical (vehicle) and transport engineering. Within this test section of the road, there is a system that simulates rain and produces artificial fog. The test rig allows rainfall simulation with intensity from 2 mm/h to 63 mm/h and reduces visibility from 90 to 3 m. As the entire section is in an area unprotected from wind and sun, the reproducibility of the results is certainly more difficult to obtain than in the case of the tests performed in the test chamber. Still, the possibilities of simulating rain here are much better than in the case of the Clermont chamber [4].

Another solution is to create a test rig from the beginning to the end, designed for a specific experiment. An example of a self-built test rig can be an external rain simulator. It is a simple structure consisting of a frame and water spray nozzles (or pipes) mounted on it. In the presented example, to increase the degree of representation of the real conditions, the researchers introduced separate controls for each of the nozzles. By gradually increasing the number and power of working nozzles, they simulated rain of varying intensity (from 12 to 120 mm/h) and different (but more precisely undefined) droplet sizes. This test rig is shown in the Fig. 2 below [9].

Fig. 2.
figure 2

Outdoor rain simulator. Source: [9].

Full size image

Such a test rig has the main advantages that it is simple and quick to construct, and at the same time, it allows to simulate rain on a similar level as a professional CERNA chamber. In addition, a small outdoor test rig gives great opportunities to change the scene and set up objects.

To sum up, for research on the operation of LiDAR sensors in adverse weather conditions, it is possible to use both professional chambers successfully, and their research infrastructure, as well as research, test rigs constructed for the needs of a specific experiment. It should be noted that the detailed methodology for both construction and simulation of weather conditions in test rigs has not been developed and rigardised so far. Individual teams of researchers individually undertake various activities aimed at ensuring the highest possible accuracy of measurements. Therefore, it is impossible to distinguish one best method of constructing such a test rig, but there are certain features that such a test rig should have:

  • the ability to ensure repeatability of measurements,

  • the ability to simulate various weather conditions (e.g. rain of varying intensity).

  • the ability to change the location of the tested device,

  • size matched to the parameters of the tested device,

  • a variety of exposure targets (reflectors with a fixed reflectivity coefficient, road signs, horizontal road signs, road infrastructure elements, mannequins imitating pedestrians, etc.),

  • possibility of implementing new illumination targets,

  • the ability to change the position of the targets for exposure.

3 Designed Test Rig

When starting the construction of the position, the goal was to fulfil all the criteria listed in the earlier part of the paper. For this purpose, it was decided to build an external test rig consisting of 3 elements: a rain shower (simulator) with a hydrophore, LiDAR sensor and reflective targets (scene).

The construction of the rain shower is made of a wooden frame equipped with four water pipes. For this purpose, composite pipes with a diameter of one quarter inch were used. Their main advantage lies in the quick and easy assembly and the possibility of attaching measuring devices, e.g., pressure gauges, to them. To ensure even distribution of water, each of the pipes was drilled at 15 points. The pipes, 150 cm long, are each arranged in parallel, with 50 cm intervals. The rig is designed so that it is possible to increase the intensity of rainfall both by filling consecutive pipes with water and by increasing or reducing the water flow separately for each pipe. The adjustment is made independently for each of the pipes. The water pressure in the system was confirmed by the indications of two independent pressure gauges permanently installed in the system. The rig was designed in such a way that its structure enables its easy modification in the event of the necessity to simulate the change of intensification of rainfall. Constructed rain shower and the reflective targets are shown in Fig. 3.

Fig. 3.
figure 3

Source: own research.

Constructed rain shower and reflective targets.

Full size image

An important issue is the need to ensure adequate water pressure. To maintain the accuracy of the measurements and repeatability of the results, it was necessary to ensure that the water was always flowing in the system under sufficiently high pressure.

Table 2. Test rig parameters, source: own research.
Full size table

For this purpose, a 750 W pump was used, and it was connected to a water tank (reservoir) with a nominal capacity of 100 L. The basic parameters of the test rig are presented in the Table 2.

The rain shower allows you to simulate heavy rain in the range of 224 l/h to 568 l/h. Its design enables to change the rain intensity in the future.

The LiDAR LIVOX Horizon was used as the measurement sensor. It is a pulsed solid-state (non-oscillating) LiDAR designed and manufactured as a sensor for use in a autonomous car. The basic parameters of the LiDAR sensor used are presented in the table below (Table 3):

Table 3. Livox Horizon Basic Specs, Source: Livox Website.
Full size table

The selected LiDAR sensor has a wide FOV, and its effective range is up to 240 m. The targets illuminated by LiDAR were models of road signs and a model of a passenger car, made on a 1:5 scale. Mockups were prepared on a scale because the effective range of LiDAR sensor is much larger than the size of the test rig. The light reflectivity of an object is shown on the screen in colours from blue with low reflectivity to red at its highest values.

Fig. 4.
figure 4

Source: own research.

Point cloud generated by LiDAR displayed in Livox Viewer.

Full size image

The mockups were made of reflective material with a very high coefficient of light reflectivity (Lambert reflection) oscillating around 255.00 (the maximum possible to obtain). As a result, the illuminated targets are shown in red in the image returned by the device. The image of the point cloud generated by the LiDAR sensor, seen in the LivoxViewer program (software provided by the LiDAR manufacturer - Livox), is shown in Fig. 4.

4 Experimental Research

As part of the experiment, rain with three intensity levels was simulated: below16.6 mm/h (<135 l/h) (drops below the effective operating range of the flow meter), 16.6 mm/h (224 l/h) and 25.8 mm/h (348 l/h). Visualisations of the obtained point cloud in the LivoxViewer program (software provided by the LiDAR manufacturer - Livox) are shown in the figures below:

Fig. 5.
figure 5

Source: own research.

Comparison of images from LiDAR: a) image without rain b) rain below 16.6 mm/h c) rain 16.6 mm/h, d) rain 25.8 mm/h.

Full size image

In order to analyse the obtained images, a detection algorithm was used. This algorithm works by recognising the colour of pixels. The first step was to read the searched pixel colour - the algorithm read the RGB colour (76,0,15). A colour selection tolerance of 10% was allowed for all components of the colours. The detection result for the Fig. 5a (without rain) is shown in Fig. 6:

Fig. 6.
figure 6

Source: own research.

Detection result for the image without rain.

Full size image

An additional filter based on a mask was used to increase detection accuracy. A mask is an algorithm that analyses groups of pixels in a given area. Without a mask, each pixel is analysed separately, while the mask filter increases the analysed area from one pixel (1x1) to the area of three pixels by three pixels (3x3). A given field is marked if at least half of its pixels meet the given colour criteria. The effect of the filtering algorithm is shown in Fig. 7.

The mask filter allowed for a more accurate marking of the searched object and eliminated the distortions previously visible at the top of the drawing. As the use of a mask filter increases detection accuracy, the analysis of the remaining images was performed using it.

Fig. 7.
figure 7

Source: own research.

Detection results for an image using a 3x3 pixel mask. a) no rain, b) rain below 16.6 mm/h, c) rain 16.6 mm/h, d) rain 25.8 mm/h.

Full size image

Figure 7a and 7b are almoust the same, while Fig. 7c shows that one of the illuminated targets has not been identified. The exact results of the detection of targets are presented in Table 4:

Table 4. Detection results for RGB (76,0,15) ± 10% color and 3x3 pixel mask filter.
Full size table

The total number of pixels in the analysed area was 7,826. For the image without rain, the number of pixels detected was the largest, and objects marked in this way constituted 20.96% of the entire image. In the case of rain below 16.6 mm/h and 16.6 mm/h, the results were very similar. There was a slight decrease in the number of pixels detected (1,461 and 1,518), while the number of objects detected did not change. The greatest decrease in the number of pixels detected was for rain with an intensity of 25.8 mm/h (1,236), which corresponds to a decrease in the number of pixels detected in relation to their total number by 5.17%. Moreover, due to the deterioration of the detection quality caused by heavy rain, one of the target objects, a mockup of a road sign, was not recognised.

5 Conclusion

The literature review proves that it is possible to build an effective, external LiDAR sensor test rig. The rigs built outside are much cheaper and easier to build than complex test chambers, and at the same time, they ensure acceptable quality and repeatability of the performed measurements.

As part of the investigation, an external test rig was built, consisting of a rain shower, LIVOX LiDAR sensor, vehicle models and road signs. The rig provides a constant water pressure of an appropriate value allowing to obtain repeatable pressure values in the pipes responsible for obtaining the required intensity of rainfall. The use of mockups of the car and road signs made in the scale of 1:5 allowed to optimise the dimensions of the rig while maintaining the required accuracy of the analysis of the obtained image. The use of the mask filtering algorithm allowed for the elimination of disturbances and an increase in the accuracy of the analysis.

The preliminary tests allowed for assessing the impact of rainfall intensity on the obtained LIDAR sensor image. The research was conducted for the rain of three different intensity levels. The analysis of materials showed that rain with the intensity (0–16.6 mm/h) is a slight disturbance to the LiDAR sensor operation, and the reduction in the number of pixels detected was from 8.5% to 11%. Neither target ceased to be visible. On the other hand, rain of greater intensity significantly disturbed the operation of the LiDAR sensor, reducing the detected pixels by 24.6%. The disruption of LIDAR sensor operation was so significant that one of the illuminated targets - a road sign mockup, was not detected.

The conducted preliminary research allows for the direction of further research on the operation of LiDAR sensor and image recognition algorithms, improving its performance in the rain.