The first step in autonomous driving is localization. Knowing its location is crucial for path planning and vehicle control. Currently, cars primarily use GPS navigation. However, GPS isn't very precise, only accurate to within 2 meters in 95% of cases. When a human is driving, knowing the approximate location is sufficient; a person can judge their position based on their surroundings. Autonomous driving, however, isn't quite that intelligent yet, so the accuracy requirements for location are much higher. If the error is too large, positioning errors may occur. For example, if you're not at an intersection but the car is located at one, it might make incorrect predictions.
What are the vehicle positioning methods used in autonomous driving systems?
GPS positioning: GPS has some problems. First, its accuracy is not high enough. Then, people use RTK technology, which involves setting up a base station with known precise coordinates (which can be measured in advance). Then, the base station and the terminal simultaneously send GPS positioning data. The deviation between the location queried by the base station and the absolute location of the base station is calculated, and the terminal uses this deviation to correct its own position. Under good signal conditions, the accuracy can reach the centimeter level, which is sufficient for most requirements.
INS positioning: As an inertial navigation sensor that is not affected by external interference but is prone to large errors due to time drift and temperature drift, it has some auxiliary functions in autonomous driving. For example, it provides predictive models for data fusion in GNSS/INS/LiDAR/HD Map fusion schemes (see Apollo's data fusion scheme, which is described in detail in a 2017 paper).
However, the accuracy of conventional INS is questionable. Apollo uses Novatel's IMU, which is moderately priced, since the IMU in this solution only provides one prediction.
LiDAR's positioning: As a sensor that can provide a large amount of external information, I think it is LiDAR that has given rise to autonomous driving, while computer vision technology has only given rise to assisted driving. Therefore, I personally have a very high level of recognition for LiDAR, and I also believe that the dozens of LiDAR manufacturers worldwide can bring down the price of LiDAR.
Velodyne's 32-line LiDAR is moderately priced and can provide millions of point clouds per frame. It can be used not only for target detection but also for point cloud matching and localization (matching real-time point clouds with high-precision maps for localization). When the high-precision map is not obstructed and there are relatively many environmental features, the localization accuracy is objective. Using GNSS/INS/LiDAR/HD Map, it can achieve the localization level of many startups or OEMs' demos.
