Among the many challenges involved in the development of safe, reliable advanced driver assist systems (ADAS), sensing and perception in adverse weather remains one of the most difficult problems.
In fact, a recent article in Bloomberg magazine entitled “Self-Driving Cars Can Handle Neither Rain nor Sleet nor Snow” claimed that “The ultimate hurdle to the next phase of driver-less technology might not come from algorithms and artificial intelligence—it might be fog and rain [1]”.
The primary technical challenges associated with automated and autonomous driving in rain come from the influence of rain on the vehicles sensors such as cameras and LIDAR. Although the qualitative impacts of weather on these sensors has been studied for quite some time [2], there has been surprisingly little progress in quantitatively predicting the impact of rain on LIDAR sensors typically used in ADAS systems.
Such a model would be useful in both defining a performance envelope for ADAS systems and in the development of weather-aware ADAS algorithms. Ideally, the model would depend on simple environment parameters such as rain rate and simple sensor parameters such as laser power. While recent measurements and a resulting empirical model of the influence of heavy rain on a Hokuyo UTM-30LX-EW were published by [3], their work measured rain rates of 40.5–95.4 mm/h, whereas naturally occurring rain rarely exceeds 25 mm/h. Therefore, the empirical model is not directly applicable to other sensors at lower rain rates. It is more useful to have a physically-based model that is relevant for a variety of realistic sensors and rain-rates.
More recently, ref. [4] published experimental results quantifying the influence of rain on the reflected intensity of a Velodyne VLP-16 sensor. Even though the empirical results are again not generally applicable to all rain rates and sensors, the results are useful in constraining the analytical model developed in this work.
Perhaps the most detailed work on the influence of rain on LIDAR sensors is [5], which gave quantitative predictions for the LIDAR range reduction as a function of rain rate and compared these to laboratory measurements. However, this model requires a detailed measurement of LIDAR specifications and in fact used a LIDAR sensor which is not currently commercially available. Because it is not often possible to easily acquire detailed internal specifications of a LIDAR sensor, a model is developed in this work that uses a simple parametrization of the LIDAR sensor to make predictions of the reflected intensity and range reduction caused by rain. The model is integrated into a physics-based simulator for autonomous driving and several simulated experiments are performed. An ADAS algorithm for obstacle detection is used to evaluate the performance reduction caused by rain in a realistic test environment.
The following sections will discuss the materials and methods used for the experiments, including a detailed description of the software (Section 2), followed by a presentation of the results of several simulated experiments (Section 3). Finally, the consequences of the results will be discussed (Section 4) followed by a brief conclusion to the paper (Section 5).
For the entire paper click here.
