Why is sensor calibration necessary for autonomous driving? Why is sensor calibration important?

Why is Sensor Calibration Necessary for Autonomous Driving?

Sensor calibration is the basis of autonomous driving perception and planning tasks. First, the perception results of each sensor need to be unified into the vehicle system fusion expression. For example, the vehicle in front perceived by Mono3D and the vehicle in front perceived by laser need to be converted to the vehicle system before they can be fused and output to the downstream. Second, some perception tasks rely on sensor extrinsic parameter calibration, such as visual IPM transformation, which requires knowing the camera extrinsic parameters. Third, front fusion tasks, such as camera and laser front fusion, also require knowing the camera & laser extrinsic parameters. Therefore, sensor calibration is the basis of all perception tasks.

Why Sensor Calibration is Important

The accuracy of sensor calibration determines the perception performance. For example, if the direction angle is off by 0.5 degrees, the lateral deviation of 100m ranging will be 100*tan(0.5 degrees) = 0.87m, which is close to 1m. Vehicles that are driving close to the lane may be misjudged as intruding the lane, causing avoidance or even emergency braking. For another example, if the direction angle of the laser and the camera deviates by 0.2 degrees, the laser point p(100, 100, 0) at 45 degrees to the side is projected by a camera with internal parameters of M【2000, 0, 960, 0, 2000, 540, 0, 0, 1】. The pixel error can be calculated as: deltaP = M*R*p - M*p, which is about 5 pixels of error. Therefore, many pre-fusion tasks cannot be done.

The application scenario determines the form of sensor calibration. The first is definitely the scenario of vehicle production and delivery. Calibration in the factory is both accurate and fast. However, those who have worked on autonomous driving know that in the early stage of research and development, there is no ideal platform support, and many things are changed while they are being done. The data collected a year ago, the car has been dismantled, and now it needs to be recalibrated. There is only a pile of data in hand, so we can only use data calibration, which is the so-called target-free calibration. For example, after the vehicle is delivered to the user, the sensor is reinstalled after repair, and it is definitely not realistic to send it back to the high-precision calibration room. In many cases, a simple calibration environment will be set up in a 4S store for calibration, and some are directly calibrated online, that is, the vehicle is calibrated on open roads according to specific requirements.

What are the autonomous driving sensor calibration tasks?

In the early stage, it is generally necessary to maintain the calibration of test vehicles and support some specific calibration requirements, such as calibration of old data without vehicles and rapid calibration of some sensors when the calibration room is incomplete. In the mid-term stage, it is generally necessary to design the calibration specification process, automate the calibration process, and develop and improve the calibration function. The late stage is mainly testing and function maintenance. The mid-term and late stages evolve from each other without absolute boundaries.

The main tasks are: offline calibration, which is the last process of the vehicle on the production line, supports the placement of some high-precision targets; online calibration, which is also called no-target calibration in some places. After the vehicle is sold, it is in the hands of the user, or old data, etc., there is no target, and only environmental information can be used for calibration. I personally think this part is the most difficult. After-sales calibration (offline calibration), after-sales maintenance scenarios, supports the placement of some simple targets and operating procedures. Of course, the same is true in the development stage. In many cases, the environment is limited, and it is very expensive to build a high-precision calibration room, but simple targets can be placed, similar to after-sales calibration, which I call offline calibration.

What can this autonomous driving sensor calibration learn?

If you want to systematically understand the calibration of autonomous driving, or if you have encountered a bottleneck and want to learn some solutions, you can refer to this course. This course first introduces the characteristics of common sensors, and then introduces the corresponding calibration solutions for common calibration requirements, focusing on the online calibration method. The course will also have a lot of code practice and homework, and you will have a deeper understanding in practice.