Application of LiDAR in driveless technology

电子科技大学格拉斯哥学院 2017级 贺俊皓
The car accident happened in Chongqing last month is really a tragedy for all peopel in China.The unreasonalbe female passenger was accused of the crime culprit of this accident. So, which kind of measure could be adapted to ensure the safety of passengers? I think driveless technology is one of the solutions.
The success of driverless cars involves sophisticated mapping, real-time positioning and obstacle detection, all of which require LiDAR. This paper will deeply analyze how lidar is widely used in various technologies of unmanned vehicle. This paper first introduces the working principle of lidar, including how to scan point cloud by laser. Then it explains the application of lidar in unmanned driving technology in detail, including mapping, positioning and obstacle detection. Finally, the challenges of lidar technology are discussed, including external environment interference, large amount of data and high cost.
Introduction to unmanned driving technology
Unmanned driving technology is the integration of many technologies, including sensors, positioning and deep learning, high-precision map, path planning, obstacle detection and avoidance, mechanical control, system integration and optimization, energy consumption and heat dissipation management and so on. Although there are many differences in the implementation of existing unmanned vehicles, the system architecture is much the same. Figure 1 shows the general system architecture of the unmanned vehicle. The sensing end of the system (left of figure 1) is composed of different sensors, of which GPS is used for positioning, Light Detection And Ranging (LiDAR) is used for positioning And obstacle Detection, And the camera is used for object recognition And positioning assistance based on deep learning.

After the acquisition of sensor information, we entered the perception stage, mainly including positioning and object recognition . In this stage, we can use mathematical methods, such as Kalman Filter, Particle Filter and other algorithms, to fuse various sensor information to obtain the current position with the maximum probability. If LiDAR is used as the main positioning sensor, the information scanned by LiDAR can be compared with the known high-precision map to get the current position of the vehicle. If there is no map, we can even compare the current LiDAR scanning information with the previous scanning information using ICP algorithm to calculate the current vehicle position. After obtaining the position prediction based on LiDAR, it can use mathematical methods to fuse with other sensor information to calculate more accurate position information.
Finally, we enter the planning and control phase (figure 1, right). At this stage, we adjust the vehicle’s driving plan in real time according to the location information and the identified image information (such as traffic lights), and convert the driving plan into a control signal to control the vehicle. The global path planning can be implemented by algorithms similar to a-star, while the local path planning can be implemented by algorithms such as DWA.
Basic knowledge of optical radar
First, let’s understand how lidar works, especially the process of generating point clouds.
The working principle of
Lidar is an optical remote sensing technology. It firstly emits a laser beam to the target object and then determines the actual distance of the target object according to the time interval between receiving and reflecting. Then, according to the distance and the Angle of laser emission, the position information of the object can be deduced by simple geometric changes. Because the propagation of laser is less affected by the outside world, the distance LiDAR can detect can generally reach more than 100m. Compared with the radio waves used by traditional radar, LiDAR USES laser rays. The wavelength of laser rays used by commercial LiDAR is generally between 600nm and 1000nm, far lower than that used by traditional radar. Therefore, LiDAR can achieve higher accuracy in measuring the distance and surface shape of objects, generally reaching the centimeter level.
LiDAR system is generally divided into three parts: the first is the laser emitter, which emits laser rays with a wavelength of 600nm to 1000nm; The second part is the scanning and optical components, which are mainly used to collect the time and Azimuth of the reflection point distance from the point. The third part is the light-sensitive part, which mainly detects the intensity of the returned light. Therefore, every point we detect includes spatial coordinate information (x, y, z) and light intensity information (I). The light intensity is directly related to the reflectivity of the object, so I can preliminarily judge the detected object according to the light intensity.
What is a point cloud?
The LiDAR used by driverless cars is not stationary. In the process of driving the unmanned vehicle, LiDAR also rotates at a certain angular speed at a constant speed. In this process, it continuously emits laser and collects information of reflection points, so as to obtain comprehensive environmental information. In the process of collecting the distance between reflection points, LiDAR also records the time and Azimuth of the reflection points at the same time, and each laser emitter has a number and a fixed vertical Angle. Based on these data, we can calculate the coordinates of all reflection points. A point cloud is formed by the collection of all reflection point coordinates collected by LiDAR every rotation.

The distance between LiDAR and the object can be measured by laser reflection. Since the vertical Angle of the laser is fixed, let’s call it a. In this case, we can directly calculate the z-axis coordinate as sin(a)*distance. We can obtain the projection of distance on xy plane by cosine (a)*distance, xy_dist. LiDAR records the distance of reflection point and also records the horizontal Angle b of the current rotation of LiDAR. According to the simple set conversion, the X-axis coordinates and Y-axis coordinates of this point are cosine (b)*xy_dist and sine (b)*xy_dist, respectively.
Application of LiDAR in unmanned driving technology
Next, how to apply lidar in unmanned driving technology is introduced, especially for the drawing of high-precision map, point cloud-based positioning and obstacle detection.
High definition map drawing
The high-resolution map here is different from the navigation map we use every day. High-definition map is a Mosaic of many point clouds, mainly used for the precise positioning of unmanned vehicles. High-definition maps are also drawn using LiDAR. The lidar-equipped map data acquisition vehicle repeatedly drives on the route where it wants to draw a high-resolution map and collects point cloud data. In the later stage, some error information in the point cloud map was filtered by manual annotation, such as points formed by reflections of cars and pedestrians on the road, and then the point cloud collected for many times was aligned and stitching to form the final high-definition map.
Location based on point cloud
First, introduce the importance of positioning. Many people have such a question: if you have accurate GPS, do not know the current location, also need to locate it? Not really. At present, the military differential GPS with high precision can achieve the accuracy of centimeter level in the “ideal” environment when it is static. The “ideal” environment here is one in which there is no excess suspended media in the atmosphere and GPS has a strong reception signal when measuring. However, unmanned vehicles drive in a complex dynamic environment, especially in large cities. Due to the obstruction of various tall buildings, the problem of GPS multi-path reflection will be more obvious. In this way, the GPS positioning information obtained is easy to have errors of tens of centimeters or even meters. For cars traveling at high speed on a limited width, such errors are likely to lead to traffic accidents. Therefore, there must be means other than GPS to enhance the accuracy of unmanned vehicle positioning.
As mentioned above, LiDAR constantly collects point clouds to understand the surrounding environment while driving. We can naturally think of using this environmental information to locate. Here, we can express this problem as a simplified probability problem: given GPS information at time t0, point cloud information at time t0, and three possible locations of unmanned vehicle at time t1: P1, P2 and P3 (for the purpose of simplifying the problem, let’s assume that unmanned vehicle will be in one of these three locations). So let’s figure out the probability that at time t1 we have a car at any of these three points. According to bayes’ law, the positioning problem of unmanned vehicle can be simplified into the following probability formula:
Picture description
The first term on the right represents the probability distribution of observed point cloud information given the current position. The calculation method is generally divided into two kinds: sub-bureau estimation and global estimation. The simple method of local estimation is to estimate the possibility of unmanned vehicle at the current position by matching point cloud at the current moment with point cloud at the previous moment and by means of geometric derivation. Global estimation is to use the point cloud at the current moment to match the hd map mentioned above, so as to obtain the possibility of the current car relative to a certain position on the map. In practice, two positioning methods are generally used together. The second term on the right represents the probability distribution of the current position prediction, where the location information given by GPS can be simply used as the prediction. By calculating the posterior probability of P1, P2 and P3, it is possible to estimate which position the unmanned vehicle is most likely to be in. By multiplying the two probability distributions, the positioning accuracy of unmanned vehicle can be greatly improved.

Although driverless technology is maturing, LiDAR is still a hurdle. Although the price of pure vision, GPS/IMU positioning and obstacle avoidance schemes is low, they are not mature enough to be applied to outdoor scenes. But at the same time, LiDAR prices remain high, making it difficult for consumers to afford driverless cars priced in the hundreds of thousands of dollars. Therefore, it is imperative to quickly reduce the cost of the system, especially the cost of LiDAR, significantly. One promising approach is to use a lower-priced LiDAR, which loses some accuracy, but can be used with other lower-priced sensors to mix information with the LiDAR to more accurately calculate the vehicle’s position. In other words, it is to make up for the deficiency of hardware sensors through better algorithms, which we believe is the development direction of unmanned vehicles in the near future. The price of high-precision LiDAR will also decline in the next year or two due to the increased market demand, paving the way for the further popularization of unmanned vehicles.