Why sensor calibration for autonomous driving?
Sensor calibration is the basis for the perception & planning task of autonomous driving. First, the results of each sensor's perception need to be unified into the vehicle system for fusion expression, for example, the front vehicle perceived by Mono3D and the front vehicle perceived by laser need to be converted to the vehicle system before they can be fused and output to the downstream. Second, some perceptual tasks rely on sensor external parameter calibration, such as visual IPM transformation, which requires knowledge of the camera's external parameters. Third, pre-fusion tasks, such as camera and laser pre-fusion, also need to know the external parameters of the camera & laser. Therefore, sensor calibration is the basis of all perception tasks.
Why sensor calibration is important
The accuracy of the sensor calibration determines the perceptual performance. For example, if the direction angle is deviated by 0.5 degrees, the 100m ranging will deviate by 100*tan (0.5 degrees) = 0.87m, close to 1m, and the vehicle that may be running close to the lane will be misjudged as invading the lane, causing avoidance or even sudden braking action. For example, the laser with a laser and camera direction angle deviation of 0.2 degrees, and the laser spot p(100, 100, 0) with a side of 45 degrees Under the camera with an internal reference of M[2000, 0, 960, 0, 2000, 540, 0, 0, 1], the projection pixel error can be calculated: deltaP = M*R*p - M*p, there is an error of about 5 pixels, then many pre-fusion things cannot be done.
The application scenario determines the calibration form of the sensor. First of all, it must be the scene of vehicle production and delivery, and the calibration in the factory is accurate and fast. However, those who have done autonomous driving know that in the early stage of research and development, there is no ideal platform support, and many things are changed while doing it, and the data collected a year ago, the car has long been dismantled, and now it is necessary to recalibrate, and there is only a bunch of data in hand, so you can only use data calibration, that is, the so-called standardless target calibration. Another example is the delivery of the user's vehicle, after the repair of the sensor reinstalled, and then sent back to the high-precision calibration room is certainly unrealistic, many times will be in the 4S shop to build a simple calibration environment for calibration, there are also some direct online calibration, that is, the vehicle in accordance with specific requirements in the open road exercise, calibration.
What are the tasks of calibrating sensors for autonomous driving?
In the initial stages, it is generally necessary to maintain the calibration of the test vehicle to support some specific calibration needs, such as calibration without the old data of the vehicle, and the rapid calibration of some sensors when the calibration room is not perfect. In the mid-term, the calibration specification process is generally designed, the calibration process is automated, and the calibration function is developed and improved. The later stage is mainly testing and functional maintenance, and the middle and late stages are mutually evolving, and there is no absolute boundary.
The main tasks are: off-line calibration, the last process of the vehicle on the production line, support the layout of some high-precision targets; online calibration, in some places also called no-target calibration, the vehicle is sold in the hands of the user, or ancient data, etc., there is no target, only the use of environmental information calibration, I personally think this part is the most difficult. After-sales calibration (offline calibration), after-sales maintenance scenarios, support the layout of some simple targets and operation processes, of course, in the development stage is the same, many times the environment is limited, it is very expensive to build a high-precision calibration room, but you can arrange simple targets, similar to after-sales calibration, I collectively refer to offline calibration.
What can be learned from this autonomous driving sensor calibration
So if you want to systematically understand the calibration of autonomous driving, or if you have encountered a bottleneck and want to understand some solutions, you can refer to this course. This course first introduces the characteristics of commonly used sensors, and then introduces the corresponding calibration schemes for common calibration requirements, focusing on online calibration methods. The course will also have a lot of code practice and assignments, and you will experience it more deeply in practice.
Course Q&A
The Q&A of this course is mainly answered in the corresponding goose circle of this course, and students can ask questions in the goose circle at any time if they have any questions during the learning process.
▲Add a small assistant: cv3d007, consult more