Unmanned is not a single point of technology, but a number of technical integration. Unmanned overall technology architecture, probably can be divided into three modules: algorithms, systems and cloud platform. In the vehicle side, the upper layer is the algorithm module, including sensing, perception Sensor and decision-making three parts; the lower is the operating system and hardware platform. In the cloud, there is an unmanned cloud platform, which includes a high-precision map, model training, simulation and data storage and other pieces of content.
Unmanned car driving process is the most important thing to know their location, we usually use the multi-star GPS, you can receive Galileo or the Big Dipper signal, do a comprehensive, and then come to a relatively accurate location. But the positioning accuracy of multi-satellite GPS can do about 1-2 meters, which can not meet the requirements of unmanned driving lane-level positioning. Follow-up, the industry has developed a known as the RTK (the GPS, mainly rely on the ground in the base station, through the base station signal to correct Fuel Rail Pressure Sensor the satellite signal, the accuracy can reach the decimeter level or even lower. But the disadvantage is the need to lay the base station , The cost is very high, but also people need to maintain.Then, there is a called PPP (precision single point positioning) GPS technology, based on the global satellite network system, through the Internet to send satellite correction signal, the technology probably 2018 years will be deployed in the world. The advantage is that there is no need to lay the base station, no matter where, can get a more accurate location.
Combine GPS and IMU inertial navigation systems. Inertial navigation system can provide rapid update - 1000 frames / second, which makes up for the low rate of GPS update; and inertial navigation exists in the 'cumulative error' problem, you can also make up for the GPS. The benefits of laser radar is a certain range, can reach 100-200 meters distance, can be very accurate to get the point in space. The laser radar data and high precision map data to do a match, you can move the vehicle positioning to centimeter level. Lidar is very dependent on another sensor - high-precision map (the traditional sense can not be called the sensor), the two can be used in order to achieve a good positioning effect. In the bottom Temperature Sensor of the high-precision map, is a grid map, the grid map is scanned using a laser radar, the accuracy of up to 5 cm; grid map above, we will do the road label, that is, the bottom of the reference Add some semantic information to the lane, mark the lanes; above the lane, do some semantic tags, such as speed limits, traffic lights such signs.
The traditional approach is to use binocular vision navigation, left and right two figures come in, first do a triangular imaging, you can draw the depth of the space information, each feature points are described, and then before and after the two images Feature points are compared to get the information of its displacement, can be positioned to the distance of the vehicle to move. The development of follow-up technology to achieve the monocular vision for navigation function, but the image information update rate is limited - between 30-60 frames / sec. So, in order to get information updates faster, or to add IMU, it produced a Visual inertial odometry technology, you can get a very accurate location update. But in practice, only rely on a positioning technology or sensor, obviously Speed Sensor can not achieve good results. Weather, light, magnetic field, etc., will interfere with the normal use of these positioning sensors.
This architecture is the mainstream unmanned companies are using, such as Stanford University's unmanned vehicles, CMU unmanned vehicles, Google's unmanned vehicles and Baidu's unmanned vehicles, are such a structure.