Specific applications of lidar in autonomous driving

LiDAR, short for Light Detection and Ranging, is used to acquire data and generate precise digital models. Its fundamental operating principle is no different from that of radio radar; a signal is emitted by the radar transmission system, reflected by the target, and collected by the receiving system. The target's distance is determined by measuring the travel time of the reflected light. However, compared to other radars, LiDAR has unique advantages, such as high resolution, strong resistance to active interference, small size, and light weight. Compared to detection tools such as cameras, its greatest advantage is its ability to generate three-dimensional positional information, quickly determining the position, size, external shape, and even material of an object, while simultaneously acquiring data and generating precise digital models.

1. Positioning

Positioning is crucial in autonomous driving. Only with real-time positional information can the system make the next judgment, deciding where to go and how to get there.
There are many ways to achieve positioning. For example, Real-Time Kinematic (RTK) carrier-phase differential technology, but RTK is still susceptible to signal interference. Especially in areas with many buildings and trees, or when entering and exiting tunnels, its signal is easily interrupted. There are also methods that use cameras and other sensors to perceive the external environment, build an environmental model, and use this model to determine the vehicle's position. However, this method is heavily reliant on the environment; for example, in backlight or rain/snow conditions, this positioning method is prone to failure. LiDAR, on the other hand, relies on comparing the vehicle's initial position with high-precision map information to obtain a precise position. First, sensors such as GPS, IMU, and wheel speed provide an initial (approximate) position. Second, local point cloud information from the LiDAR undergoes feature extraction, and combined with the initial position, vector features in the global coordinate system are obtained. Finally, the vector features from the previous step are matched with the feature information of the high-precision map to obtain precise global positioning. Therefore, in terms of both accuracy and stability, LiDAR offers unparalleled advantages in positioning. Its only drawback is the currently high production cost, but I believe that cost reduction is an inevitable trend. On the one hand, costs can be reduced through large-scale production; on the other hand, technological innovation in the direction of solid-state LiDAR is underway, aiming for low-cost and mass-producible solid-state LiDAR. Many domestic and foreign manufacturers are accelerating innovation, and in the near future, cost will not be a major problem.

2. Obstacle Detection and Classification
For obstacle detection and classification, both vision and LiDAR are currently used, and these two methods are not mutually exclusive. LiDAR is not dependent on lighting; its field of view is 360 degrees, the computational load is relatively small, and it can scan in real time (currently, generally within 100 milliseconds). During the scanning process, LiDAR first identifies obstacles and determines their spatial positions. Then, it classifies the existing obstacles. For example, cars and people. We segment these obstacles into independent entities and match them to classify obstacles and track objects. The tracking process first involves segmenting the point cloud, associating points to targets, determining whether the previous and next frames belong to the same object, and then performing object tracking to output object tracking information.

3. Used in Advanced Driver-Assistance Systems (ADAS)
Advanced Driver-Assistance Systems (ADAS) utilize various sensors installed in the vehicle to collect environmental data inside and outside the vehicle in real time. This data undergoes processing, including the identification, detection, and tracking of static and dynamic objects. This allows drivers to quickly perceive potential dangers, increasing awareness and improving safety through active safety technology.
If LiDAR can effectively control costs, even lower-level ADAS driving assistance functions will require LiDAR. This is because camera-based ADAS and autonomous driving systems, or those using only millimeter-wave radar, have significant limitations. First, there is the issue of the field of view. To ensure sufficient detection distance, the field of view angle cannot be too large, which results in a very large lateral blind spot for the vehicle. Some companies have designed multiple cameras to address this issue, but the results are not ideal. Multiple cameras have overlapping areas and increase the difficulty of data processing. Second, there is the low-speed problem. In fact, at low speeds, cameras perform poorly and struggle to identify even slowly moving or stationary targets. Therefore, LiDAR has great potential in ADAS.