Challenges
AV applications primarily rely on a set of sensors and camera systems. Adverse weather conditions such as fog, heavy rain, snow, and wind can severely limit the functionality of sensors and cameras. Although the applications appear to work well in dry, sunny weather, those are just the best-case scenarios.
We know that the adverse is a tricky situation to handle for AV sensors. Is it worth AV developer's attention and how much effort to put on it? and how much improvement it could bring for AV drive safety to a higher level? Which sensor perform better in the worse weather scenarios? To start answer the questions, let's search for the statistics...
Car Accident (according to weather condition)
Take a example from France metropolitan record of year 2020*The statistics showcase the car accidents mostly happening (approx. 80%) is at the normal and covered weather condition. Among those adverse weather, more accidents (10%) happened during the light rain weather condition.
*Resource: French Road Safety Observatory (ONISR)
Sensors
Among the camera, LIDAR, RADAR and ultrasonic sensors, RADAR performs very good at detecting moving objects under fog, rain, snow, and dust. But it cannot work alone to enhance safety applications. Detecting of static and recognition of objects and road feathers still need to be performed by vision detectors like camera.
Performance Degradation
The foggy air blurs also the camera image and decreases the quality and counts of the lidar point cloud. From the hands-on work experience I have, more degradation happens on the cover window of cameras and lidar. This phenomena could be modeled then simulated, and compared with the laboratory or even road drive test results. This practice of modeling and test correlation study is very helpful to understand the driving cases.
Degradation Study Method...
Analysis and Simulation
Of course data gathering of different road scenarios are mandatory step for data preparation, learning and data training process. I would highlight the importance of properly understanding the weather condition, which plays the key role for building mathematical models, in order to further anticipate each sensor's behavior under certain conditions. In this way the proper model would help estimate the degradation of sensor performance, and adjust data sets, filter out noise brought in by the nature of rain drops.
Physical Modeling...
Possible Solution...
Comentários